Can you post code snippet? Pastbin link is fine. -Vlad
On Tue, Aug 11, 2015 at 4:03 PM, Serega Sheypak <serega.shey...@gmail.com> wrote: > Probably I found something. Response time decreases when parallelism grows. > What I did: > > 1. wrap business logic controller into java main class. My controller does > some logic and puts/gets to hbase with checkAndPut (sometimes) > 2. create HConnection > 3. pass HConnection to controller > 4. wrap controller execution into codahale metrics > 5. execute controller in several threads simultaneously. The same happens > in servlet environment > > I can't explain result. > 1. I used 10 threads and 100000 iterations in each. > > *RESULT: 99% <= 28.81 milliseconds which sounds GOOD!* > -- Meters > ---------------------------------------------------------------------- > putMeter > count = 414914 > mean rate = 885.58 events/second > 1-minute rate = 911.56 events/second > 5-minute rate = 778.16 events/second > 15-minute rate = 549.72 events/second > > -- Timers > ---------------------------------------------------------------------- > putTimer > count = 414914 > mean rate = 884.66 calls/second > 1-minute rate = 911.53 calls/second > 5-minute rate = 765.60 calls/second > 15-minute rate = 515.06 calls/second > min = 4.87 milliseconds > max = 211.77 milliseconds > mean = 10.81 milliseconds > stddev = 5.43 milliseconds > median = 10.34 milliseconds > 75% <= 11.59 milliseconds > 95% <= 14.41 milliseconds > 98% <= 19.59 milliseconds > 99% <= 28.81 milliseconds > 99.9% <= 60.67 milliseconds > > I've increased count of threads to 100: > *RESULT: 99% <= 112.09 milliseconds* > -- Meters > ---------------------------------------------------------------------- > putMeter > count = 1433056 > mean rate = 2476.46 events/second > 1-minute rate = 2471.18 events/second > 5-minute rate = 2483.28 events/second > 15-minute rate = 2512.52 events/second > > -- Timers > ---------------------------------------------------------------------- > putTimer > count = 1433058 > mean rate = 2474.61 calls/second > 1-minute rate = 2468.45 calls/second > 5-minute rate = 2446.45 calls/second > 15-minute rate = 2383.23 calls/second > min = 10.03 milliseconds > max = 853.05 milliseconds > mean = 40.71 milliseconds > stddev = 39.04 milliseconds > median = 35.60 milliseconds > 75% <= 47.69 milliseconds > 95% <= 71.79 milliseconds > 98% <= 85.83 milliseconds > 99% <= 112.09 milliseconds > 99.9% <= 853.05 milliseconds > > Is it possible to explain it? Could it be a problem in some > pooling/threading inside HConnection? > > please see what happened to compactions during test: > http://www.bigdatapath.com/wp-content/uploads/2015/08/compations.png > > get/put ops > http://www.bigdatapath.com/wp-content/uploads/2015/08/get_ops.png > > slow ops: > http://www.bigdatapath.com/wp-content/uploads/2015/08/slow_ops.png > > 2015-08-11 23:43 GMT+02:00 Serega Sheypak <serega.shey...@gmail.com>: > > > >How about GC activity? ApplicationStopTime? Do you track that? > > yes, jviusalm says it's ok, newrelic also doesn't show something strange. > > HBase also says it's OK. > > > > Profiler says most time thread is waiting for response from hbase side. > My > > assumption is: > > 1. I have weird bug in HBase configuration > > 2. I have undiscovered problems with networking (BUT the same tomcats > > write data to flume with higher rate, no data loss at all) > > 3. I have weird problem with HConnection HConnectionManager is > > multithreaded env, when same servlet instance shared across many threads > > 4. some mystic process somewhere in the cluster.... > > > > >Is the issue reproducible? or you got it first time? > > always. Spikes disappear during night, but RPM doesn't change too much. > > > > I will run my controller code out of tomcat and see how it goes. I'm > going > > to isolate components... > > > > > > 2015-08-11 23:36 GMT+02:00 Vladimir Rodionov <vladrodio...@gmail.com>: > > > >> How about GC activity? ApplicationStopTime? Do you track that? > >> > >> Is the issue reproducible? or you got it first time? > >> > >> Start with RS logs and try to find anything suspicious in a period of a > >> very high latency. 1.5 sec HBase write latency does not look right. > >> > >> -Vlad > >> > >> On Tue, Aug 11, 2015 at 2:08 PM, Serega Sheypak < > serega.shey...@gmail.com > >> > > >> wrote: > >> > >> > Hi Vladimir! > >> > > >> > Here are graphs. Servlet (3 tomcats on 3 different hosts write to > HBase) > >> > http://www.bigdatapath.com/wp-content/uploads/2015/08/01_apps1.png > >> > See how response time jump. I can't explain it. Write load is > >> really-really > >> > low. > >> > > >> > all RS have even load. I see request-metrics in HBase master web UI. > >> > Tables are pre-splitted. I have 10 RS and pre-splitted tables on 50 > >> > regions. > >> > > >> > >1. How large is your single write? > >> > 1-2KB > >> > > >> > >2. Do you see any RegionTooBusyException in a client log files > >> > no HBase related exceptions. Response > >> > > >> > > 3. How large is your table ( # of regions, # of column families) > >> > 1 column familiy, table_01 150GB, table_02 130 GB > >> > > >> > I have two "major tables", here are stats for them: > >> > http://www.bigdatapath.com/wp-content/uploads/2015/08/table_02.png > >> > http://www.bigdatapath.com/wp-content/uploads/2015/08/table_01.png > >> > >4. RS memory related config: Max heap > >> > > >> > 5. memstore size (if not default - 0.4) > >> > hbase.regionserver.global.memstore.upperLimit=0.4 > >> > hbase.regionserver.global.memstore.lowerLimit=0.38 > >> > RS heapsize=8GB > >> > > >> > >*Do you see any region splits? * > >> > no, never happened since tables are pre-splitted > >> > > >> > 2015-08-11 18:54 GMT+02:00 Vladimir Rodionov <vladrodio...@gmail.com > >: > >> > > >> > > *Common questions:* > >> > > > >> > > > >> > > 1. How large is your single write? > >> > > 2. Do you see any RegionTooBusyException in a client log files > >> > > 3. How large is your table ( # of regions, # of column families) > >> > > 4. RS memory related config: Max heap > >> > > 5. memstore size (if not default - 0.4) > >> > > > >> > > > >> > > Memstore flush > >> > > > >> > > hbase.hregion.memstore.flush.size = 256M > >> > > hbase.hregion.memstore.block.multiplier = N (do not block writes) N > * > >> > 256M > >> > > MUST be greater than overall memstore size (HBASE_HEAPSIZE * > >> > > hbase.regionserver.global.memstore.size) > >> > > > >> > > WAL files. > >> > > > >> > > Set HDFS block size to 256MB. hbase.regionserver.hlog.blocksize = > 0.95 > >> > HDFS > >> > > block size (256MB * 0.95). Keep hbase.regionserver.hlog.blocksize * > >> > > hbase.regionserver.maxlogs just a bit above > >> > > hbase.regionserver.global.memstore.lowerLimit > >> > > (0.35-0.45) * HBASE_HEAPSIZE to avoid premature memstore flushing. > >> > > > >> > > *Do you see any region splits? * > >> > > > >> > > Region split blocks writes. Try to presplit table and avoid > splitting > >> > after > >> > > that. Disable splitting completely > >> > > > >> > > hbase.regionserver.region.split.policy > >> > > =org.apache.hadoop.hbase.regionserver.DisabledRegionSplitPolicy > >> > > > >> > > -Vlad > >> > > > >> > > > >> > > > >> > > > >> > > On Tue, Aug 11, 2015 at 3:22 AM, Serega Sheypak < > >> > serega.shey...@gmail.com> > >> > > wrote: > >> > > > >> > > > Hi, we are using version 1.0.0+cdh5.4.4+160 > >> > > > We have heavy write load, ~ 10K per econd > >> > > > We have 10 nodes 7 disks each. I read some perf notes, they state > >> that > >> > > > HBase can handle 1K per second writes per node without any > problems. > >> > > > > >> > > > > >> > > > I see some spikes on "writers". Write operation timing "jumps" > from > >> > > 40-50ms > >> > > > to 200-500ms Probably I hit memstore limit. RegionServer starts to > >> > flush > >> > > > memstore and stop to accept updates. > >> > > > > >> > > > I have several questions: > >> > > > 1. Does 4/(8 in hyperthreading) CPU + 7HDD node could absorb 1K > >> writes > >> > > per > >> > > > second? > >> > > > 2. What is the right way to fight with blocked writes? > >> > > > 2.1. What I did: > >> > > > hbase.hregion.memstore.flush.size to 256M to produce larger HFiles > >> when > >> > > > flushing memstore > >> > > > base.hregion.memstore.block.multiplier to 4, since I have only one > >> > > > intensive-write table. Let it grow > >> > > > hbase.regionserver.optionallogflushinterval to 10s, i CAN loose > some > >> > > data, > >> > > > NP here. The idea that I reduce I/O pressure on disks. > >> > > > === > >> > > > Not sure if I can correctly play with these parameters. > >> > > > hbase.hstore.blockingStoreFiles=10 > >> > > > hbase.hstore.compactionThreshold=3 > >> > > > > >> > > > >> > > >> > > > > >
