Thanks for sharing this, Matt. Do you mind opening a Jira for your suggestion ?
On Feb 25, 2012, at 3:18 PM, Matt Corgan <[email protected]> wrote: > I've been meaning to look into something regarding compactions for a while > now that may be relevant here. It could be that this is already how it > works, but just to be sure I'll spell out my suspicions... > > I did a lot of large uploads when we moved to .92. Our biggest dataset is > time series data (partitioned 16 ways with a row prefix). The actual > inserting and flushing went extremely quickly, and the parallel compactions > were churning away. However, when the compactions inevitably started > falling behind I noticed a potential problem. The compaction queue would > get up to, say, 40, which represented, say, an hour's worth of requests. > The problem was that by the time a compaction request started executing, > the CompactionSelection that it held was terribly out of date. It was > compacting a small selection (3-5) of the 50 files that were now there. > Then the next request would compact another (3-5), etc, etc, until the > queue was empty. It would have been much better if a CompactionRequest > decided what files to compact when it got to the head of the queue. Then > it could see that there are now 50 files needing compacting and to possibly > compact the 30 smallest ones, not just 5. When the insertions were done > after many hours, I would have preferred it to do one giant major > compaction, but it sat there and worked through it's compaction queue > compacting all sorts of different combinations of files. > > Said differently, it looks like .92 picks the files to compact at > compaction request time rather than compaction execution time which is > problematic when these times grow far apart. Is that the case? Maybe > there are some other effects that are mitigating it... > > Matt > > On Sat, Feb 25, 2012 at 10:05 AM, Jean-Daniel Cryans > <[email protected]>wrote: > >> Hey guys, >> >> So in HBASE-4365 I ran multiple uploads and the latest one I reported >> was a 5TB import on 14 RS and it took 18h with Stack's patch. Now one >> thing we can see is that apart from some splitting, there's a lot of >> compacting going on. Stack was wondering exactly how much that IO >> costs us, so we devised a test where we could upload 5TB with 0 >> compactions. Here are the results: >> >> The table was pre-split with 14 regions, 1 per region server. >> hbase.hstore.compactionThreshold=100 >> hbase.hstore.blockingStoreFiles=110 >> hbase.regionserver.maxlogs=64 (the block size is 128MB) >> hfile.block.cache.size=0.05 >> hbase.regionserver.global.memstore.lowerLimit=0.40 >> hbase.regionserver.global.memstore.upperLimit=0.74 >> export HBASE_REGIONSERVER_OPTS="$HBASE_JMX_BASE -Xmx14G >> -XX:CMSInitiatingOccupancyFraction=75 -XX:NewSize=256m >> -XX:MaxNewSize=256m" >> >> The table had: >> MAX_FILESIZE => '549755813888', MEMSTORE_FLUSHSIZE => '549755813888' >> >> Basically what I'm trying to do is to never block and almost always be >> flushing. You'll probably notice the big difference between the lower >> and upper barriers and think "le hell?", it's because it takes so long >> to flush that you have to have enough room to take on more data while >> this is happening (and we are able to flush faster than we take on >> write). >> >> The test reports the following: >> Wall time: 34984.083 s >> Aggregate Throughput: 156893.07 queries/s >> Aggregate Throughput: 160030935.29 bytes/s >> >> That's 2x faster than when we wait for compactions and splits, not too >> bad but I'm pretty sure we can do better: >> >> - The QPS was very uneven, it seems that when it's flushing it takes >> a big toll and queries drop to ~100k/s while the rest of the time it's >> more like 200k/s. Need to figure out what's going there and if it's >> really just caused by flush-related IO. >> - The logs were rolling every 6 seconds and since this takes a global >> write lock, I can see how we could be slowing down a lot across 14 >> machines. >> - The load was a bit uneven, I miscalculated my split points and the >> last region always had 2-3k more queries per second. >> >> Stay tuned for more. >> >> J-D >>
