Hi Lars, We are following your suggestion and testing against a single region server. We just ran a test against a remote region server and soon we will test against a local one as well. We will get back to you soon with the results.
It will take us a couple of days to port to and test our code with 0.94.2. Once we have it working, we will run some experiments and update this thread. Unfortunately, the nature of our project is such that we cannot easily translate the benchmark's workload into a program executing the equivalent HBase operations directly. For this reason, I attempted to roughly translate the workload in terms of HBase operations in my first email and I attached a portion of the logs to be a bit more concrete. Your assistance is very much appreciated! Thank you! We'll keep you updated. Best regards, Yousuf On Fri, Oct 19, 2012 at 1:25 AM, lars hofhansl <lhofha...@yahoo.com> wrote: > Can you reproduce this against a single, local region server? > Any chance that you can try with the just released 0.94.2? > > > I would love to debug this. If would be a tremendous help if you had a > little test program that reproduces this against a single server, so that I > can see what is going on. > > Thanks. > > -- Lars > > > > ----- Original Message ----- > From: Yousuf Ahmad <myahm...@gmail.com> > To: user@hbase.apache.org; lars hofhansl <lhofha...@yahoo.com> > Cc: Ivan Brondino <ibrond...@fi.upm.es>; Ricardo Vilaça < > rmvil...@di.uminho.pt> > Sent: Thursday, October 18, 2012 12:59 PM > Subject: Re: High IPC Latency > > Hi, > > Thank you for your questions guys. > > We are using HBase 0.92 with HDFS 1.0.1. > > The experiment lasts 15 minutes. The measurements stabilize in the first > two minutes of the run. > > The data is distributed almost evenly across the regionservers so each > client hits most of them over the course of the experiment. However, for > the data we have, any given multi-get or scan should touch only one or at > most two regions. > > The client caches the locations of the regionservers, so after a couple of > minutes of the experiment running, it wouldn't need to re-visit ZooKeeper, > I believe. Correct me if I am wrong please. > > Regards, > Yousuf > > > On Thu, Oct 18, 2012 at 2:42 PM, lars hofhansl <lhofha...@yahoo.com> > wrote: > > > Also, what version of HBase/HDFS is this using? > > > > > > > > > > ----- Original Message ----- > > From: "Pamecha, Abhishek" <apame...@x.com> > > To: "user@hbase.apache.org" <user@hbase.apache.org> > > Cc: Ivan Brondino <ibrond...@fi.upm.es>; Ricardo Vilaça < > > rmvil...@di.uminho.pt> > > Sent: Thursday, October 18, 2012 11:38 AM > > Subject: RE: High IPC Latency > > > > Is it sustained for the same client hitting the same region server OR > does > > it get better for the same client-RS combination when run for longer > > duration? Trying to eliminate Zookeeper from this. > > > > Thanks, > > Abhishek > > > > From: Yousuf Ahmad [mailto:myahm...@gmail.com] > > Sent: Thursday, October 18, 2012 11:26 AM > > To: user@hbase.apache.org > > Cc: Ivan Brondino; Ricardo Vilaça > > Subject: High IPC Latency > > > > Hello, > > > > We are seeing slow times for read operations in our experiments. We are > > hoping that you guys can help us figure out what's going wrong. > > > > Here are some details: > > > > * We are running a read-only benchmark on our HBase cluster. > > * > > * There are 10 regionservers, each co-located with a datanode. HDFS > > replication is 3x. > > * All the data read by the experiment is already in the block cache > > and the hit ratio is 99%. > > * > > * We have 10 clients, each with around 400 threads making a mix of > > read-only requests involving multi-gets and scans. > > * > > * We settled on the default client pool type/size (roundrobin/1) and > a > > regionserver handler count of 100 after testing various combinations to > see > > what setting worked best. > > * > > * Our scans are short, fetching around 10 rows on average. Scanner > > caching is set to 50. > > * An average row in a scan has either around 10 columns (small row) > or > > around 200 columns (big row). > > * > > * Our multi-gets fetch around 200 rows on average. > > * An average row in a multi-get has around 10 columns. > > * Each column holds an integer (encoded into bytes). > > * > > * None of the machines involved reach CPU, memory, or IO saturation. > > In fact resource utilization stays quite low. > > * > > * Our statistics show that the average time for a scan, measured > > starting from the first scanner.next() call to the last one which > returns a > > null, is around 2-3 seconds. > > * Since we use scanner caching, the major portion of this time > (around > > 2 seconds) is spent on the first call to next(), while the remaining > calls > > take a negligible amount of time. > > * Similarly, we see that a multi-get on average takes around 2 > seconds. > > * A single get on average takes around 1 second. > > We are not sure what the bottleneck is or where it lies. We thought we > > should look deeper into what is going on at the regionservers. We > monitored > > the IPC calls during one of the experiments. Here is a sample of one > > regionserver log: > > > > 2012-10-18 17:00:09,969 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115483; Served: HRegionInterface#get queueTime=0 processingTime=1 > > contents=1 Get, 75 bytes > > 2012-10-18 17:00:09,969 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115487; Served: HRegionInterface#get queueTime=0 processingTime=0 > > contents=1 Get, 75 bytes > > 2012-10-18 17:00:09,969 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115489; Served: HRegionInterface#get queueTime=0 processingTime=0 > > contents=1 Get, 75 bytes > > 2012-10-18 17:00:09,982 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #111421; Served: HRegionInterface#get queueTime=0 processingTime=0 > > contents=1 Get, 75 bytes > > 2012-10-18 17:00:09,982 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115497; Served: HRegionInterface#multi queueTime=0 processingTime=9 > > contents=200 Gets > > 2012-10-18 17:00:09,984 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115499; Served: HRegionInterface#openScanner queueTime=0 > > processingTime=0 contents=1 Scan, 63 bytes > > 2012-10-18 17:00:09,990 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #115503; Served: HRegionInterface#get queueTime=0 processingTime=0 > > contents=1 Get, 75 bytes > > 2012-10-18 17:00:09,992 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #103230; Served: HRegionInterface#next queueTime=0 processingTime=0 > > contents=1 Long, 1 Integer > > 2012-10-18 17:00:09,994 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #103234; Served: HRegionInterface#close queueTime=0 processingTime=0 > > contents=1 Long > > 2012-10-18 17:00:09,994 DEBUG org.apache.hadoop.ipc.HBaseServer.trace: > > Call #103232; Served: HRegionInterface#next queueTime=0 processingTime=0 > > contents=1 Long, 1 Integer > > > > I have attached a larger chunk of the logs we collected for this > > experiment in case that helps. > > > > From the logs, we saw that the next() operation at the regionserver takes > > 1 millisecond or less; and a multi-get takes 10 ms on average. > > Yet the corresponding times we see at the client are orders of magnitude > > higher. > > Ping times between the machines are at most 1ms and we are not saturating > > the network. > > > > We would really appreciate some insights from you guys on this. > > Where do you suggest we focus our efforts in order to hunt down this > > bottleneck/contention? > > > > Thanks! > > Yousuf > > > >
