400mb blockcache? Ouch. What's your hbase-env.sh? Have you configured a heap size? My guess is you're using the un configured default of 1G. Should be at least 8G, and maybe more like 30G with this kind of host.
How many users are sharing it and with what kinds of tasks? If there's no IO isolation between processes, I suspect your benchmarks will be worthless on this shared environment. -n On Friday, November 21, 2014, Liu, Ming (HPIT-GADSC) <ming.l...@hp.com> wrote: > Thank you Lars, > > There must be something wrong with my testing yesterday. I cannot > reproduce the issue anymore. Now, changing the cache.size from 0 to 0.4 > will not slow down the write perf dramatically, but still will slow down > write (Yesterday I saw a 7x slowdown, today it is about 1.3x slowdown which > is acceptable for me). As Ted pointed out, it is possible that memstore > cannot get enough memory when more RAM give to block cache so it flush more > frequently, but I really need more reading to understand when memstore will > flush. > And one thing I noticed is when I restart hbase for the very first test, > the performance is best, then the second time, it is slower, both read and > write, and slower and slower in the following tests and get to a stable > point after about 3 or 4 times, in each run I will read 5,000,000 rows and > update 5,000,000 rows. There are too many factors affect the read/write OPS > in hbase... > > My purpose is to find a proper way to evaluate performance, since we are > going to change something in hbase and it is good to have a base benchmark > so we can compare the performance after change. So I must make sure the > perf test itself make sense and should be trusted. > > I saw an entry in the log may help to see the cache settings in my system: > hfile.LruBlockCache: Total=373.54 MB, free=13.13 MB, max=386.68 MB, > blocks=5655, accesses=17939947, hits=14065015, hitRatio=78.40%, , > cachingAccesses=17934857, cachingHits=14064420, cachingHitsRatio=78.42%, > evictions=15646, evicted=3861343, evictedPerRun=246.7942657470703 > > My testing environment is a workstation with 12 core CPU, 96G memory and > 1.7 T disk. But it is a shared workstation, many users share it and I > started hbase in standalone mode with hbase-site.xml as below : > > <configuration> > <property> > <name>hbase.rootdir</name> > <value>hdfs://localhost:24400/hbase</value> > </property> > <property> > <name>hbase.zookeeper.property.dataDir</name> > <value>hdfs://localhost:24400/zookeeper</value> > </property> > <property> > <name>hbase.master.port</name> > <value>24560</value> > </property> > <property> > <name>hbase.master.info.port</name> > <value>24561</value> > </property> > <property> > <name>hbase.regionserver.port</name> > <value>24562</value> > </property> > <property> > <name>hbase.regionserver.info.port</name> > <value>24563</value> > </property> > <property> > <name>hbase.zookeeper.peerport</name> > <value>24567</value> > </property> > <property> > <name>hbase.zookeeper.leaderport</name> > <value>24568</value> > </property> > <property> > <name>hbase.zookeeper.property.clientPort</name> > <value>24570</value> > </property> > <property> > <name>hbase.rest.port</name> > <value>24571</value> > </property> > <property> > <name>hbase.client.scanner.caching</name> > <value>100</value> > </property> > <property> > <name>hbase.client.scanner.timeout.period</name> > <value>60000</value> > </property> > <property> > <name>hbase.bulkload.staging.dir</name> > <value>hdfs://localhost:24400/hbase-staging</value> > </property> > <property> > <name>hbase.snapshot.enabled</name> > <value>true</value> > </property> > <property> > <name>hbase.master.distributed.log.splitting</name> > <value>false</value> > </property> > <property> > <name>zookeeper.session.timeout</name> > <value>90000000</value> :-) I just want to make sure never timeout > here, since I get timeout so many times... > </property> > <property> > <name>hfile.block.cache.size</name> > <value>0.4</value> > </property> > </configuration> > > > [liuliumi@ YCSB]$ free -m > total used free shared buffers cached > Mem: 96731 46828 49903 0 984 32525 > -/+ buffers/cache: 13317 83413 > Swap: 48432 12 48420 > > > [liuliumi@ YCSB]$ lscpu > Architecture: x86_64 > CPU op-mode(s): 32-bit, 64-bit > Byte Order: Little Endian > CPU(s): 12 > On-line CPU(s) list: 0-11 > Thread(s) per core: 1 > Core(s) per socket: 6 > Socket(s): 2 > NUMA node(s): 2 > Vendor ID: GenuineIntel > CPU family: 6 > Model: 44 > Stepping: 2 > CPU MHz: 3065.894 > BogoMIPS: 6130.96 > Virtualization: VT-x > L1d cache: 32K > L1i cache: 32K > L2 cache: 256K > L3 cache: 12288K > NUMA node0 CPU(s): 0,2,4,6,8,10 > NUMA node1 CPU(s): 1,3,5,7,9,11 > > > Thanks, > Ming > > -----Original Message----- > From: lars hofhansl [mailto:la...@apache.org <javascript:;>] > Sent: Friday, November 21, 2014 4:31 AM > To: user@hbase.apache.org <javascript:;> > Subject: Re: how to explain read/write performance change after modifying > the hfile.block.cache.size? > > That would explain it if memstores are flushed due to global memory > pressure. > > But cache and memstore size are (unfortunately) configured independently. > The memstore heap portion would be 40% (by default) in either case.So this > is a bit curious still. > Ming, can you tell us more details?- RAM on the boxes- heap setup for the > region servers- any other relevant settings on hbase-site.xml- configs on > the table/column family you're writing to (like bloom filters, etc). > > That would help us diagnose this. > > -- Lars > > From: Ted Yu <yuzhih...@gmail.com <javascript:;>> > To: "user@hbase.apache.org <javascript:;>" <user@hbase.apache.org > <javascript:;>> > Sent: Thursday, November 20, 2014 9:32 AM > Subject: Re: how to explain read/write performance change after modifying > the hfile.block.cache.size? > > When block cache size increases from 0 to 0.4, the amount of heap given to > memstore decreases. This would slow down the writes. > Please see: > http://hbase.apache.org/book.html#store.memstore > > For your second question, see this thread: > > http://search-hadoop.com/m/DHED4TEvBy1/lars+hbase+hflush&subj=Re+Clarifications+on+HBase+Durability > > Cheers > > On Thu, Nov 20, 2014 at 8:05 AM, Liu, Ming (HPIT-GADSC) <ming.l...@hp.com > <javascript:;>> > wrote: > > > Hello, all, > > > > I am playing with YCSB to test HBase performance. I am using HBase > 0.98.5. > > I tried to adjust the hfile.block.cache.size to see the difference, > > when I set hfile.block.cache.size to 0, read performance is very bad, > > but write performance is very very very good....; when I set > > hfile.block.cache.size to 0.4, read is better, but write performance > > drop dramatically. I disable the client side writebuffer already. > > This is hard to understand for me: > > The HBase guide just said hfile.block.cache.size setting is about how > > much memory used as block cache used by StoreFile. I have no idea of > > how HBase works internally. Typically, it is easy to understand that > > increase the size of cache should help the read, but why it will harm > > the write operation? The write performance down from 30,000 to 4,000 > > for your reference, just by changing the hfile.block.cache.size from 0 > to 0.4. > > Could anyone give me a brief explanation about this observation or > > give me some advices about what to study to understand what is block > cache used for? > > > > Another question: HBase write will first write to WAL then to memstore. > > Will the write to WAL go to disk directly before hbase write memstore, > > a sync operation or it is possible that write to WAL is still buffered > > somewhere when hbase put the data into the memstore? > > > > Reading src code may cost me months, so a kindly reply will help me a > > lot... ... > > Thanks very much! > > > > Best Regards, > > Ming > > > > > >
