Gotcha. So if I'm using java maps then my tserver_opts needs to be tserver.memory.maps + extra for the rest of the tserver because the memory map will be taken from the overall memory allocated to the tserver. But if I'm using native maps, then I need far less tserver memory because the map memory is not deducted from the tserver. Is that correct?
Thanks, tejay From: John Vines [mailto:[email protected]] Sent: Friday, September 21, 2012 8:26 AM To: [email protected] Subject: Re: EXTERNAL: Re: Failing Tablet Servers memory.maps is what defines the size of the in memory map. When using native maps, that space does not come out of the heap size. But when using non-native maps, it comes out of the heap space. I think the issue Eric is trying to hit at is the fickleness of the java garbage collector. When you give a process that much heap, that's so much more data you can hold before you need to garbage collect. However, that also means when it does garbage collect, it's collecting a LOT more, which can result is poor performance. John On Fri, Sep 21, 2012 at 10:12 AM, Cardon, Tejay E <[email protected]<mailto:[email protected]>> wrote: Jim, Eric, and Adam, Thanks. It sounds like you're all saying the same thing. Originally I was doing each key/value as its own mutation, and it was blowing up much faster (probably due to the volume/overhead of the mutation objects themselves. I'll try refactoring to break them up into something in-between. My keys are small (<25 Bytes), and my values are empty, but I'll aim for ~1,000 key/values per mutation and see how that works out for me. Eric, I was under the impression that the memory.maps setting was not very important when using native maps. Apparently I'm mistaken there. What does this setting control when in a native map setting? And, in general, what's the proper balance between tserver_opts and tserver.memory.maps? With regards to the "Finished gathering information from 24 servers in 27.45 seconds" Do you have any recommendations for how to chase down the bottleneck? I'm pretty sure I'm having GC issues, but I'm not sure what is causing them on the server side. I'm sending a fairly small number of very large mutation objects, which I'd expect to be a moderate problem for the GC, but not a huge one.. Thanks again to everyone for being so responsive and helpful. Tejay Cardon From: Eric Newton [mailto:[email protected]<mailto:[email protected]>] Sent: Friday, September 21, 2012 8:03 AM To: [email protected]<mailto:[email protected]> Subject: EXTERNAL: Re: Failing Tablet Servers A few items noted from your logs: tserver.memory.maps.max = 1G If you are giving your processes 10G, you might want to make the map larger, say 6G, and then reduce the JVM by 6G. Write-Ahead Log recovery complete for rz<;zw== (8 mutations applied, 8000000 entries created) You are creating rows with 1M columns. This is ok, but you might want to write them out more incrementally. WARN : Running low on memory That's pretty self-explanatory. I'm guessing that the very large mutations are causing the tablet servers to run out of memory before they are held waiting for minor compactions. Finished gathering information from 24 servers in 27.45 seconds Something is running slow, probably due to GC thrashing. WARN : Lost servers [10.1.24.69:9997[139d46130344b98]] And there's a server crashing, probably due to an OOM condition. Send smaller mutations. Maybe keep it to 200K column updates. You can still have 1M wide rows, just send 5 mutations. -Eric On Thu, Sep 20, 2012 at 5:05 PM, Cardon, Tejay E <[email protected]<mailto:[email protected]>> wrote: I'm seeing some strange behavior on a moderate (30 node) cluster. I've got 27 tablet servers on large dell servers with 30GB of memory each. I've set the TServer_OPTS to give them each 10G of memory. I'm running an ingest process that uses AccumuloInputFormat in a MapReduce job to write 1,000 rows with each row containing ~1,000,000 columns in 160,000 families. The MapReduce initially runs quite quickly and I can see the ingest rate peak on the monitor page. However, after about 30 seconds of high ingest, the ingest falls to 0. It then stalls out and my map task are eventually killed. In the end, the map/reduce fails and I usually end up with between 3 and 7 of my Tservers dead. Inspecting the tserver.err logs shows nothing, even on the nodes that fail. The tserver.out log shows a java OutOfMemoryError, and nothing else. I've included a zip with the logs from one of the failed tservers and a second one with the logs from the master. Other than the out of memory, I'm not seeing anything that stands out to me. If I reduce the data size to only 100,000 columns, rather than 1,000,000, the process takes about 4 minutes and completes without incident. Am I just ingesting too quickly? Thanks, Tejay Cardon
