Thanks Lars. I have changed the BLOCKSIZE to 16KB and triggered a major compaction. I will report my results once it is done.
- Ramu On Mon, Oct 7, 2013 at 3:21 PM, lars hofhansl <la...@apache.org> wrote: > First of: 128gb heap per RegionServer. Wow.I'd be interested to hear your > experience with such a large heap for your RS. It's definitely big enough. > > > It's interesting hat 100gb do fit into the aggregate cache (of 8x32gb), > while 1.8tb do not. > Looks like ~70% of the read request would need to bring in a 64kb block in > order to read 724 bytes. > > Should that take 100ms? No. Something's still amiss. > > Smaller blocks might help (you'd need to bring in 4, 8, or maybe 16k to > read the small row). You would need to issue a major compaction for that to > take effect. > Maybe try 16k blocks. If that speeds up your random gets we know where to > look next... At the disk IO. > > > -- Lars > > > > ________________________________ > From: Ramu M S <ramu.ma...@gmail.com> > To: user@hbase.apache.org; lars hofhansl <la...@apache.org> > Sent: Sunday, October 6, 2013 11:05 PM > Subject: Re: HBase Random Read latency > 100ms > > > Lars, > > In one of your old posts, you had mentioned that lowering the BLOCKSIZE is > good for random reads (of course with increased size for Block Indexes). > > Post is at http://grokbase.com/t/hbase/user/11bat80x7m/row-get-very-slow > > Will that help in my tests? Should I give it a try? If I alter my table, > should I trigger a major compaction again for this to take effect? > > Thanks, > Ramu > > > > On Mon, Oct 7, 2013 at 2:44 PM, Ramu M S <ramu.ma...@gmail.com> wrote: > > > Sorry BLOCKSIZE was wrong in my earlier post, it is the default 64 KB. > > > > {NAME => 'usertable', FAMILIES => [{NAME => 'cf', DATA_BLOCK_ENCODING => > > 'NONE', BLOOMFILTER => 'ROWCOL', REPLICATION_SCOPE => '0', VERSIONS => > '1', > > COMPRESSION => 'NONE', MIN_VERSIONS => '0', TTL => '2147483647', > > KEEP_DELETED_CELLS => 'false', BLOCKSIZE => '65536', IN_MEMORY => > 'false', > > ENCODE_ON_DISK => 'true', BLOCKCACHE => 'true'}]} > > > > Thanks, > > Ramu > > > > > > On Mon, Oct 7, 2013 at 2:42 PM, Ramu M S <ramu.ma...@gmail.com> wrote: > > > >> Lars, > >> > >> - Yes Short Circuit reading is enabled on both HDFS and HBase. > >> - I had issued Major compaction after table is loaded. > >> - Region Servers have max heap set as 128 GB. Block Cache Size is 0.25 > of > >> heap (So 32 GB for each Region Server) Do we need even more? > >> - Decreasing HFile Size (Default is 1GB )? Should I leave it to default? > >> - Keys are Zipfian distributed (By YCSB) > >> > >> Bharath, > >> > >> Bloom Filters are enabled. Here is my table details, > >> {NAME => 'usertable', FAMILIES => [{NAME => 'cf', DATA_BLOCK_ENCODING => > >> 'NONE', BLOOMFILTER => 'ROWCOL', REPLICATION_SCOPE => '0', VERSIONS => > '1', > >> COMPRESSION => 'NONE', MIN_VERSIONS => '0', TTL => '2147483647', > >> KEEP_DELETED_CELLS => 'false', BLOCKSIZE => '16384', IN_MEMORY => > 'false', > >> ENCODE_ON_DISK => 'true', BLOCKCACHE => 'true'}]} > >> > >> When the data size is around 100GB (100 Million records), then the > >> latency is very good. I am getting a throughput of around 300K OPS. > >> In both cases (100 GB and 1.8 TB) Ganglia stats show that Disk reads are > >> around 50-60 MB/s throughout the read cycle. > >> > >> Thanks, > >> Ramu > >> > >> > >> On Mon, Oct 7, 2013 at 2:21 PM, lars hofhansl <la...@apache.org> wrote: > >> > >>> Have you enabled short circuit reading? See here: > >>> http://hbase.apache.org/book/perf.hdfs.html > >>> > >>> How's your data locality (shown on the RegionServer UI page). > >>> > >>> > >>> How much memory are you giving your RegionServers? > >>> If you reads are truly random and the data set does not fit into the > >>> aggregate cache, you'll be dominated by the disk and network. > >>> Each read would need to bring in a 64k (default) HFile block. If short > >>> circuit reading is not enabled you'll get two or three context > switches. > >>> > >>> So I would try: > >>> 1. Enable short circuit reading > >>> 2. Increase the block cache size per RegionServer > >>> 3. Decrease the HFile block size > >>> 4. Make sure your data is local (if it is not, issue a major > compaction). > >>> > >>> > >>> -- Lars > >>> > >>> > >>> > >>> ________________________________ > >>> From: Ramu M S <ramu.ma...@gmail.com> > >>> To: user@hbase.apache.org > >>> Sent: Sunday, October 6, 2013 10:01 PM > >>> Subject: HBase Random Read latency > 100ms > >>> > >>> > >>> Hi All, > >>> > >>> My HBase cluster has 8 Region Servers (CDH 4.4.0, HBase 0.94.6). > >>> > >>> Each Region Server is with the following configuration, > >>> 16 Core CPU, 192 GB RAM, 800 GB SATA (7200 RPM) Disk > >>> (Unfortunately configured with RAID 1, can't change this as the > Machines > >>> are leased temporarily for a month). > >>> > >>> I am running YCSB benchmark tests on HBase and currently inserting > around > >>> 1.8 Billion records. > >>> (1 Key + 7 Fields of 100 Bytes = 724 Bytes per record) > >>> > >>> Currently I am getting a write throughput of around 100K OPS, but > random > >>> reads are very very slow, all gets have more than 100ms or more > latency. > >>> > >>> I have changed the following default configuration, > >>> 1. HFile Size: 16GB > >>> 2. HDFS Block Size: 512 MB > >>> > >>> Total Data size is around 1.8 TB (Excluding the replicas). > >>> My Table is split into 128 Regions (No pre-splitting used, started > with 1 > >>> and grew to 128 over the insertion time) > >>> > >>> Taking some inputs from earlier discussions I have done the following > >>> changes to disable Nagle (In both Client and Server hbase-site.xml, > >>> hdfs-site.xml) > >>> > >>> <property> > >>> <name>hbase.ipc.client.tcpnodelay</name> > >>> <value>true</value> > >>> </property> > >>> > >>> <property> > >>> <name>ipc.server.tcpnodelay</name> > >>> <value>true</value> > >>> </property> > >>> > >>> Ganglia stats shows large CPU IO wait (>30% during reads). > >>> > >>> I agree that disk configuration is not ideal for Hadoop cluster, but as > >>> told earlier it can't change for now. > >>> I feel the latency is way beyond any reported results so far. > >>> > >>> Any pointers on what can be wrong? > >>> > >>> Thanks, > >>> Ramu > >>> > >> > >> > > >