For the reasons I gave above... the puts are sometimes blocked on the memstores which are blocked by the flusher thread which is blocked because there's too many files to compact because the compactor is given too many small files to compact and has to compact the same data a bunch of times.
Also, I may not have been clear, but HBase doesn't compress data in memory. This means, like I showed, that the 64MB that lives in memory becomes a 6MB file after compression (using GZ). You pack a lot more data into the same region, but performance is achieved by being able to write big files (else we wouldn't be waiting to get to 64MB before flushing). Without compression the files are much bigger and don't need as much compaction since there's so few of them, and then it splits early, to the expense of IO. If we were able to compress directly in memory and output the file as is, then we would be able to carry a lot more data in the MemStores and flush bigger files to disk... but it's not the case. Todd Lipcon once described this situation as HBase basically saying "ok you can put as fast as you can... oh wait stop stop stop that's too much... ok you can start inserting again... oh wait no that's too much" etc etc. HBase could do a better job at slowing down inserts when detecting this situation (or something like that), BTW this jira has been opened to track this issue https://issues.apache.org/jira/browse/HBASE-2981 J-D On Mon, Mar 14, 2011 at 7:04 PM, Bryan Keller <brya...@gmail.com> wrote: > I changed the settings as described below: > > hbase.hstore.blockingStoreFiles=20 > hbase.hregion.memstore.block.multiplier=4 > MAX_FILESIZE=512mb > MEMSTORE_FLUSHSIZE=128mb > > I also created the table with 6 regions initially. Before I wasn't creating > any regions initially. I needed to make all of these changes together to > entirely eliminate the very long pauses. Now there are no pauses much longer > than a second. > > Thanks much for the help. I am still not entirely sure why compression seems > to expose this problem, however. > > > On Mar 14, 2011, at 11:54 AM, Jean-Daniel Cryans wrote: > >> Alright so here's a preliminary report: >> >> - No compression is stable for me too, short pauses. >> - LZO gave me no problems either, generally faster than no compression. >> - GZ initially gave me weird results, but I quickly saw that I forgot >> to copy over the native libs from the hadoop folder so my logs were >> full of: >> >> 2011-03-14 10:20:29,624 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,626 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,628 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,630 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,632 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,634 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> 2011-03-14 10:20:29,636 INFO org.apache.hadoop.io.compress.CodecPool: >> Got brand-new compressor >> >> I copied the libs over, bounced the region servers, and the >> performance was much more stable until a point where I got a 20 >> seconds pause, and looking at the logs I see: >> >> 2011-03-14 10:31:17,625 WARN >> org.apache.hadoop.hbase.regionserver.MemStoreFlusher: Region >> test,,1300127266461.9d0eb095b77716c22cd5c78bb503c744. has too many >> store files; delaying flush up to 90000ms >> >> (our config sets the block at 20 store files instead of the default >> which is around 12 IIRC) >> >> Quickly followed by a bunch of: >> >> 2011-03-14 10:31:26,757 INFO >> org.apache.hadoop.hbase.regionserver.HRegion: Blocking updates for >> 'IPC Server handler 20 on 60020' on region >> test,,1300127266461.9d0eb095b77716c22cd5c78bb503c744.: memstore size >> 285.6m is >= than blocking 256.0m size >> >> (our settings make it that we won't block on memstores until 4x their >> sizes, in your case you may see a 2x blocking factor so 128MB which is >> default) >> >> The reason is that our memstores, once flushed, occupy a very small >> space, consider this: >> >> 2011-03-14 10:31:16,606 INFO >> org.apache.hadoop.hbase.regionserver.Store: Added >> hdfs://sv2borg169:9000/hbase/test/9d0eb095b77716c22cd5c78bb503c744/test/420552941380451032, >> entries=216000, sequenceid=70556635737, memsize=64.3m, filesize=6.0m >> >> It means that it will create tiny files of ~6MB and the compactor will >> spend all it's time merging those files until a point where HBase must >> stop inserting in order to not blow its available memory. Thus, the >> same data will get rewritten a couple of times. >> >> Normally, and by that I mean a system where you're not just trying to >> insert data ASAP but where most of your workload is made up of reads, >> this works well as the memstores are filled much more slowly and >> compactions happen at a normal pace. >> >> If you search around the interwebs for tips on speeding up HBase >> inserts, you'll often see the configs I referred to earlier: >> >> <name>hbase.hstore.blockingStoreFiles</name> >> <value>20</value> >> and >> <name>hbase.hregion.memstore.block.multiplier</name> >> <value>4</value> >> >> They should work pretty well for most use cases that are made of heavy >> writes given that the region servers have enough heap (eg more than 3 >> or 4GB). You should also consider setting MAX_FILESIZE to >1GB to >> limit the number of regions and MEMSTORE_FLUSHSIZE to >128MB to flush >> bigger files. >> >> Hope this helps, >> >> J-D >> >> On Mon, Mar 14, 2011 at 10:29 AM, Jean-Daniel Cryans >> <jdcry...@apache.org> wrote: >>> Thanks for the report Bryan, I'll try your little program against one >>> of our 0.90.1 cluster that has similar hardware. >>> >>> J-D >>> >>> On Sun, Mar 13, 2011 at 1:48 PM, Bryan Keller <brya...@gmail.com> wrote: >>>> If interested, I wrote a small program that demonstrates the problem >>>> (http://vancameron.net/HBaseInsert.zip). It uses Gradle, so you'll need >>>> that. To run, enter "gradle run". >>>> >>>> On Mar 13, 2011, at 12:14 AM, Bryan Keller wrote: >>>> >>>>> I am using the Java client API to write 10,000 rows with about 6000 >>>>> columns each, via 8 threads making multiple calls to the >>>>> HTable.put(List<Put>) method. I start with an empty table with one column >>>>> family and no regions pre-created. >>>>> >>>>> With compression turned off, I am seeing very stable performance. At the >>>>> start there are a couple of 10-20sec pauses where all insert threads are >>>>> blocked during a region split. Subsequent splits do not cause all of the >>>>> threads to block, presumably because there are more regions so no one >>>>> region split blocks all inserts. GCs for HBase during the insert is not a >>>>> major problem (6k/55sec). >>>>> >>>>> When using either LZO or gzip compression, however, I am seeing frequent >>>>> and long pauses, sometimes around 20 sec but often over 80 seconds in my >>>>> test. During these pauses all 8 of the threads writing to HBase are >>>>> blocked. The pauses happen throughout the insert process. GCs are higher >>>>> in HBase when using compression (60k, 4min), but it doesn't seem enough >>>>> to explain these pauses. Overall performance obviously suffers >>>>> dramatically as a result (about 2x slower). >>>>> >>>>> I have tested this in different configurations (single node, 4 nodes) >>>>> with the same result. I'm using HBase 0.90.1 (CDH3B4), Sun/Oracle Java >>>>> 1.6.0_24, CentOS 5.5, Hadoop LZO 0.4.10 from Cloudera. Machines have 12 >>>>> cores and 24 gb of RAM. Settings are pretty much default, nothing out of >>>>> the ordinary. I tried playing around with region handler count and >>>>> memstore settings, but these had no effect. >>>>> >>>> >>>> >>> > >
