Re: Write performance with 1.2.12
With a single node I get 3K for cassandra 1.0.12 and 1.2.12. So I suspect there is some network chatter. I have started looking at the sources, hoping to find something. 1.2 is pretty stable, I doubt there is anything in there that makes it run slower than 1.0. It’s probably something in your configuration or network. Compare the local write time from nodetool cfhistograms and the request latency from nodetool proxyhistograms. Writes request latency should be a bit below 1ms and local write latency should be around .5 ms or better. if there is a wider difference between the two it’s wait time + network time. As a general rule you should get around 3k to 4k writes per second per core. Cheers - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 13/12/2013, at 8:06 pm, Rahul Menon ra...@apigee.com wrote: Quote from http://www.datastax.com/dev/blog/performance-improvements-in-cassandra-1-2 Murmur3Partitioner is NOT compatible with RandomPartitioner, so if you’re upgrading and using the new cassandra.yaml file, be sure to change the partitioner back to RandomPartitioner On Thu, Dec 12, 2013 at 10:57 PM, srmore comom...@gmail.com wrote: On Thu, Dec 12, 2013 at 11:15 AM, J. Ryan Earl o...@jryanearl.us wrote: Why did you switch to RandomPartitioner away from Murmur3Partitioner? Have you tried with Murmur3? # partitioner: org.apache.cassandra.dht.Murmur3Partitioner partitioner: org.apache.cassandra.dht.RandomPartitioner Since I am comparing between the two versions I am keeping all the settings same. I see Murmur3Partitioner has some performance improvement but then switching back to RandomPartitioner should not cause performance to tank, right ? or am I missing something ? Also, is there an easier way to update the data from RandomPartitioner to Murmur3 ? (upgradesstable ?) On Fri, Dec 6, 2013 at 10:36 AM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00
Re: Write performance with 1.2.12
On Wed, Dec 11, 2013 at 10:49 PM, Aaron Morton aa...@thelastpickle.comwrote: It is the write latency, read latency is ok. Interestingly the latency is low when there is one node. When I join other nodes the latency drops about 1/3. To be specific, when I start sending traffic to the other nodes the latency for all the nodes increases, if I stop traffic to other nodes the latency drops again, I checked, this is not node specific it happens to any node. Is this the local write latency or the cluster wide write request latency ? This is a cluster wide write latency. What sort of numbers are you seeing ? I have a custom application that writes data to the cassandra node, so the numbers might be different than the standard stress test but it should be good enough for comparison. With the previous release 1.0.12 I was getting around 10K requests/ sec and with 1.2.12 I am getting around 6K requests/ sec. Everything else is the same. This is a three node cluster. With a single node I get 3K for cassandra 1.0.12 and 1.2.12. So I suspect there is some network chatter. I have started looking at the sources, hoping to find something. -sandeep Cheers - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 12/12/2013, at 3:39 pm, srmore comom...@gmail.com wrote: Thanks Aaron On Wed, Dec 11, 2013 at 8:15 PM, Aaron Morton aa...@thelastpickle.comwrote: Changed memtable_total_space_in_mb to 1024 still no luck. Reducing memtable_total_space_in_mb will increase the frequency of flushing to disk, which will create more for compaction to do and result in increased IO. You should return it to the default. You are right, had to revert it back to default. when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. What are you measuring, request latency or local read/write latency ? If it’s write latency it’s probably GC, if it’s read is probably IO or data model. It is the write latency, read latency is ok. Interestingly the latency is low when there is one node. When I join other nodes the latency drops about 1/3. To be specific, when I start sending traffic to the other nodes the latency for all the nodes increases, if I stop traffic to other nodes the latency drops again, I checked, this is not node specific it happens to any node. I don't see any GC activity in logs. Tried to control the compaction by reducing the number of threads, did not help much. Hope that helps. - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 7/12/2013, at 8:05 am, srmore comom...@gmail.com wrote: Changed memtable_total_space_in_mb to 1024 still no luck. On Fri, Dec 6, 2013 at 11:05 AM, Vicky Kak vicky@gmail.com wrote: Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.comwrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with
Re: Write performance with 1.2.12
Why did you switch to RandomPartitioner away from Murmur3Partitioner? Have you tried with Murmur3? 1. # partitioner: org.apache.cassandra.dht.Murmur3Partitioner 2. partitioner: org.apache.cassandra.dht.RandomPartitioner On Fri, Dec 6, 2013 at 10:36 AM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
On Thu, Dec 12, 2013 at 11:15 AM, J. Ryan Earl o...@jryanearl.us wrote: Why did you switch to RandomPartitioner away from Murmur3Partitioner? Have you tried with Murmur3? 1. # partitioner: org.apache.cassandra.dht.Murmur3Partitioner 2. partitioner: org.apache.cassandra.dht.RandomPartitioner Since I am comparing between the two versions I am keeping all the settings same. I see Murmur3Partitioner has some performance improvement but then switching back to RandomPartitioner should not cause performance to tank, right ? or am I missing something ? Also, is there an easier way to update the data from RandomPartitioner to Murmur3 ? (upgradesstable ?) On Fri, Dec 6, 2013 at 10:36 AM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Quote from http://www.datastax.com/dev/blog/performance-improvements-in-cassandra-1-2 *Murmur3Partitioner is NOT compatible with RandomPartitioner, so if you’re upgrading and using the new cassandra.yaml file, be sure to change the partitioner back to RandomPartitioner* On Thu, Dec 12, 2013 at 10:57 PM, srmore comom...@gmail.com wrote: On Thu, Dec 12, 2013 at 11:15 AM, J. Ryan Earl o...@jryanearl.us wrote: Why did you switch to RandomPartitioner away from Murmur3Partitioner? Have you tried with Murmur3? 1. # partitioner: org.apache.cassandra.dht.Murmur3Partitioner 2. partitioner: org.apache.cassandra.dht.RandomPartitioner Since I am comparing between the two versions I am keeping all the settings same. I see Murmur3Partitioner has some performance improvement but then switching back to RandomPartitioner should not cause performance to tank, right ? or am I missing something ? Also, is there an easier way to update the data from RandomPartitioner to Murmur3 ? (upgradesstable ?) On Fri, Dec 6, 2013 at 10:36 AM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Changed memtable_total_space_in_mb to 1024 still no luck. Reducing memtable_total_space_in_mb will increase the frequency of flushing to disk, which will create more for compaction to do and result in increased IO. You should return it to the default. when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. What are you measuring, request latency or local read/write latency ? If it’s write latency it’s probably GC, if it’s read is probably IO or data model. Hope that helps. - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 7/12/2013, at 8:05 am, srmore comom...@gmail.com wrote: Changed memtable_total_space_in_mb to 1024 still no luck. On Fri, Dec 6, 2013 at 11:05 AM, Vicky Kak vicky@gmail.com wrote: Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ?
Re: Write performance with 1.2.12
Thanks Aaron On Wed, Dec 11, 2013 at 8:15 PM, Aaron Morton aa...@thelastpickle.comwrote: Changed memtable_total_space_in_mb to 1024 still no luck. Reducing memtable_total_space_in_mb will increase the frequency of flushing to disk, which will create more for compaction to do and result in increased IO. You should return it to the default. You are right, had to revert it back to default. when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. What are you measuring, request latency or local read/write latency ? If it’s write latency it’s probably GC, if it’s read is probably IO or data model. It is the write latency, read latency is ok. Interestingly the latency is low when there is one node. When I join other nodes the latency drops about 1/3. To be specific, when I start sending traffic to the other nodes the latency for all the nodes increases, if I stop traffic to other nodes the latency drops again, I checked, this is not node specific it happens to any node. I don't see any GC activity in logs. Tried to control the compaction by reducing the number of threads, did not help much. Hope that helps. - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 7/12/2013, at 8:05 am, srmore comom...@gmail.com wrote: Changed memtable_total_space_in_mb to 1024 still no luck. On Fri, Dec 6, 2013 at 11:05 AM, Vicky Kak vicky@gmail.com wrote: Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80
Re: Write performance with 1.2.12
It is the write latency, read latency is ok. Interestingly the latency is low when there is one node. When I join other nodes the latency drops about 1/3. To be specific, when I start sending traffic to the other nodes the latency for all the nodes increases, if I stop traffic to other nodes the latency drops again, I checked, this is not node specific it happens to any node. Is this the local write latency or the cluster wide write request latency ? What sort of numbers are you seeing ? Cheers - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 12/12/2013, at 3:39 pm, srmore comom...@gmail.com wrote: Thanks Aaron On Wed, Dec 11, 2013 at 8:15 PM, Aaron Morton aa...@thelastpickle.com wrote: Changed memtable_total_space_in_mb to 1024 still no luck. Reducing memtable_total_space_in_mb will increase the frequency of flushing to disk, which will create more for compaction to do and result in increased IO. You should return it to the default. You are right, had to revert it back to default. when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. What are you measuring, request latency or local read/write latency ? If it’s write latency it’s probably GC, if it’s read is probably IO or data model. It is the write latency, read latency is ok. Interestingly the latency is low when there is one node. When I join other nodes the latency drops about 1/3. To be specific, when I start sending traffic to the other nodes the latency for all the nodes increases, if I stop traffic to other nodes the latency drops again, I checked, this is not node specific it happens to any node. I don't see any GC activity in logs. Tried to control the compaction by reducing the number of threads, did not help much. Hope that helps. - Aaron Morton New Zealand @aaronmorton Co-Founder Principal Consultant Apache Cassandra Consulting http://www.thelastpickle.com On 7/12/2013, at 8:05 am, srmore comom...@gmail.com wrote: Changed memtable_total_space_in_mb to 1024 still no luck. On Fri, Dec 6, 2013 at 11:05 AM, Vicky Kak vicky@gmail.com wrote: Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00
Write performance with 1.2.12
We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Hard to say much without knowing about the cassandra configurations. Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.0058.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Hi srmore, Perhaps if you use jconsole and connect to the jvm using jmx. Then uner MBeans tab, start inspecting the GC metrics. /Jason On Fri, Dec 6, 2013 at 11:40 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Looks like I am spending some time in GC. java.lang:type=GarbageCollector,name=ConcurrentMarkSweep CollectionTime = 51707; CollectionCount = 103; java.lang:type=GarbageCollector,name=ParNew CollectionTime = 466835; CollectionCount = 21315; On Fri, Dec 6, 2013 at 9:58 AM, Jason Wee peich...@gmail.com wrote: Hi srmore, Perhaps if you use jconsole and connect to the jvm using jmx. Then uner MBeans tab, start inspecting the GC metrics. /Jason On Fri, Dec 6, 2013 at 11:40 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Since how long the server had been up, hours,days,months? On Fri, Dec 6, 2013 at 10:41 PM, srmore comom...@gmail.com wrote: Looks like I am spending some time in GC. java.lang:type=GarbageCollector,name=ConcurrentMarkSweep CollectionTime = 51707; CollectionCount = 103; java.lang:type=GarbageCollector,name=ParNew CollectionTime = 466835; CollectionCount = 21315; On Fri, Dec 6, 2013 at 9:58 AM, Jason Wee peich...@gmail.com wrote: Hi srmore, Perhaps if you use jconsole and connect to the jvm using jmx. Then uner MBeans tab, start inspecting the GC metrics. /Jason On Fri, Dec 6, 2013 at 11:40 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Not long: Uptime (seconds) : 6828 Token: 56713727820156410577229101238628035242 ID : c796609a-a050-48df-bf56-bb09091376d9 Gossip active: true Thrift active: true Native Transport active: false Load : 49.71 GB Generation No: 1386344053 Uptime (seconds) : 6828 Heap Memory (MB) : 2409.71 / 8112.00 Data Center : DC Rack : RAC-1 Exceptions : 0 Key Cache: size 56154704 (bytes), capacity 104857600 (bytes), 27 hits, 155669426 requests, 0.000 recent hit rate, 14400 save period in seconds Row Cache: size 0 (bytes), capacity 0 (bytes), 0 hits, 0 requests, NaN recent hit rate, 0 save period in seconds On Fri, Dec 6, 2013 at 11:15 AM, Vicky Kak vicky@gmail.com wrote: Since how long the server had been up, hours,days,months? On Fri, Dec 6, 2013 at 10:41 PM, srmore comom...@gmail.com wrote: Looks like I am spending some time in GC. java.lang:type=GarbageCollector,name=ConcurrentMarkSweep CollectionTime = 51707; CollectionCount = 103; java.lang:type=GarbageCollector,name=ParNew CollectionTime = 466835; CollectionCount = 21315; On Fri, Dec 6, 2013 at 9:58 AM, Jason Wee peich...@gmail.com wrote: Hi srmore, Perhaps if you use jconsole and connect to the jvm using jmx. Then uner MBeans tab, start inspecting the GC metrics. /Jason On Fri, Dec 6, 2013 at 11:40 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
Re: Write performance with 1.2.12
Changed memtable_total_space_in_mb to 1024 still no luck. On Fri, Dec 6, 2013 at 11:05 AM, Vicky Kak vicky@gmail.com wrote: Can you set the memtable_total_space_in_mb value, it is defaulting to 1/3 which is 8/3 ~ 2.6 gb in capacity http://www.datastax.com/dev/blog/whats-new-in-cassandra-1-0-improved-memory-and-disk-space-management The flushing of 2.6 gb to the disk might slow the performance if frequently called, may be you have lots of write operations going on. On Fri, Dec 6, 2013 at 10:06 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:59 AM, Vicky Kak vicky@gmail.com wrote: You have passed the JVM configurations and not the cassandra configurations which is in cassandra.yaml. Apologies, was tuning JVM and that's what was in my mind. Here are the cassandra settings http://pastebin.com/uN42GgYT The spikes are not that significant in our case and we are running the cluster with 1.7 gb heap. Are these spikes causing any issue at your end? There are no big spikes, the overall performance seems to be about 40% low. On Fri, Dec 6, 2013 at 9:10 PM, srmore comom...@gmail.com wrote: On Fri, Dec 6, 2013 at 9:32 AM, Vicky Kak vicky@gmail.com wrote: Hard to say much without knowing about the cassandra configurations. The cassandra configuration is -Xms8G -Xmx8G -Xmn800m -XX:+UseParNewGC -XX:+UseConcMarkSweepGC -XX:+CMSParallelRemarkEnabled -XX:SurvivorRatio=4 -XX:MaxTenuringThreshold=2 -XX:CMSInitiatingOccupancyFraction=75 -XX:+UseCMSInitiatingOccupancyOnly Yes compactions/GC's could skipe the CPU, I had similar behavior with my setup. Were you able to get around it ? -VK On Fri, Dec 6, 2013 at 7:40 PM, srmore comom...@gmail.com wrote: We have a 3 node cluster running cassandra 1.2.12, they are pretty big machines 64G ram with 16 cores, cassandra heap is 8G. The interesting observation is that, when I send traffic to one node its performance is 2x more than when I send traffic to all the nodes. We ran 1.0.11 on the same box and we observed a slight dip but not half as seen with 1.2.12. In both the cases we were writing with LOCAL_QUORUM. Changing CL to ONE make a slight improvement but not much. The read_Repair_chance is 0.1. We see some compactions running. following is my iostat -x output, sda is the ssd (for commit log) and sdb is the spinner. avg-cpu: %user %nice %system %iowait %steal %idle 66.460.008.950.010.00 24.58 Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util sda 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sda1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sda2 0.0027.60 0.00 4.40 0.00 256.00 58.18 0.012.55 1.32 0.58 sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 sdb1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-1 0.00 0.00 0.00 0.60 0.00 4.80 8.00 0.005.33 2.67 0.16 dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-3 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.249.80 0.13 0.32 dm-4 0.00 0.00 0.00 6.60 0.0052.80 8.00 0.011.36 0.55 0.36 dm-5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.000.00 0.00 0.00 dm-6 0.00 0.00 0.00 24.80 0.00 198.40 8.00 0.29 11.60 0.13 0.32 I can see I am cpu bound here but couldn't figure out exactly what is causing it, is this caused by GC or Compaction ? I am thinking it is compaction, I see a lot of context switches and interrupts in my vmstat output. I don't see GC activity in the logs but see some compaction activity. Has anyone seen this ? or know what can be done to free up the CPU. Thanks, Sandeep
