Ok thanks Avery. But I still have two questions, the first a really dumb newbie question. Why are there ever any number of partitions other than exactly one per worker thread (one per worker in our case)? And a deeper question. Even if I shrink the cache I would suppose that if Facebook has billions of vertices they must have thousands or workers. It would seem the cache scheme simply blows up on huge graphs no matter what you do. What am I missing here?
On Wed, Sep 4, 2013 at 11:18 AM, Avery Ching <ach...@apache.org> wrote: > The amount of memory for the send message cache is per worker = > number of compute threads * number of workers * size of the cache. > > The number of partitions doesn't affect the memory usage very much. My > advice would be to dial down the cache size a bit with MAX_MSG_REQUEST_SIZE. > > Avery > > > On 9/4/13 3:33 AM, Lukas Nalezenec wrote: > > > Thanks, > I was not sure if it really works as I described. > > > Facebook can't be using it like this if, as described, they have > billions of vertices and a trillion edges. > > Yes, its strange. I guess configuration does not help so much on large > cluster. What might help are properties of input data. > > > So do you, or Avery, have any idea how you might initialize this is a > more reasonable way, and how??? > > Fast workaround is to set number of partitions to from W^2 to W or 2*W . > It will help if you dont have very large number of workers. > I would not change MAX_*_REQUEST_SIZE much since it may hurt performance. > You can do some preprocessing before loading data to Giraph. > > > > How to change Giraph: > The caches could be flushed if total sum of vertexes/edges in all caches > exceeds some number. Ideally, it should prevent not only OutOfMemory errors > but also raising high water mark. Not sure if it (preventing raising HWM) > is easy to do. > I am going to use almost-prebuild partitions. For my use case it would be > ideal to detect if some cache is abandoned and i would not be used anymore. > It would cut memory usage in caches from ~O(n^3) to ~O(n). It could be > done by counting number of cache flushes or cache insertions and if some > cache was not touched for long time it would be flushed. > > There could be separated configuration MAX_*_REQUEST_SIZE for per > partition caches during loading data. > > I guess there should be simple but efficient way how to trace memory > high-water mark. It could look like: > > Loading data: Memory high-water mark: start: 100 Gb end: 300 Gb > Iteration 1 Computation: Memory high-water mark: start: 300 Gb end: 300 Gb > Iteration 1 XYZ .... > Iteration 2 Computation: Memory high-water mark: start: 300 Gb end: 300 Gb > . > . > . > > Lukas > > > > > > On 09/04/13 01:12, Jeff Peters wrote: > > Thank you Lukas!!! That's EXACTLY the kind of model I was building in my > head over the weekend about why this might be happening, and why increasing > the number of AWS instances (and workers) does not solve the problem > without increasing each worker's VM. Surely Facebook can't be using it like > this if, as described, they have billions of vertices and a trillion edges. > So do you, or Avery, have any idea how you might initialize this is a more > reasonable way, and how??? > > > On Mon, Sep 2, 2013 at 6:08 AM, Lukas Nalezenec < > lukas.naleze...@firma.seznam.cz> wrote: > >> Hi >> >> I wasted few days on similar problem. >> >> I guess the problem was that during loading - if you have got W workers >> and W^2 partitions there are W^2 partition caches in each worker. >> Each cache can hold 10 000 vertexes by default. >> I had 26 000 000 vertexes, 60 workers -> 3600 partitions. It means that >> there can be up to 36 000 000 vertexes in caches in each worker if input >> files are random. >> Workers were assigned 450 000 vertexes but failed when they had 900 000 >> vertexes in memory. >> >> Btw: Why default number of partitions is W^2 ? >> >> (I can be wrong) >> Lukas >> >> >> >> On 08/31/13 01:54, Avery Ching wrote: >> >> Ah, the new caches. =) These make things a lot faster (bulk data >> sending), but do take up some additional memory. if you look at >> GiraphConstants, you can find ways to change the cache sizes (this will >> reduce that memory usage). >> For example, MAX_EDGE_REQUEST_SIZE will affect the size of the edge >> cache. MAX_MSG_REQUEST_SIZE will affect the size of the message cache. >> The caches are per worker, so 100 workers would require 50 MB per worker >> by default. Feel free to trim it if you like. >> >> The byte arrays for the edges are the most efficient storage possible >> (although not as performance as the native edge stores). >> >> Hope that helps, >> >> Avery >> >> On 8/29/13 4:53 PM, Jeff Peters wrote: >> >> Avery, it would seem that optimizations to Giraph have, unfortunately, >> turned the majority of the heap into "dark matter". The two snapshots are >> at unknown points in a superstep but I waited for several supersteps so >> that the activity had more or less stabilized. About the only thing >> comparable between the two snapshots are the vertexes, 192561 X >> "RecsVertex" in the new version and 191995 X "Coloring" in the old system. >> But with the new Giraph 672710176 out of 824886184 bytes are stored as >> primitive byte arrays. That's probably indicative of some very fine >> performance optimization work, but it makes it extremely difficult to know >> what's really out there, and why. I did notice that a number of caches have >> appeared that did not exist before, >> namely SendEdgeCache, SendPartitionCache, SendMessageCache >> and SendMutationsCache. >> >> Could any of those account for a larger per-worker footprint in a >> modern Giraph? Should I simply assume that I need to force AWS to configure >> its EMR Hadoop so that each instance has fewer map tasks but with a >> somewhat larger VM max, say 3GB instead of 2GB? >> >> >> On Wed, Aug 28, 2013 at 4:57 PM, Avery Ching <ach...@apache.org> wrote: >> >>> Try dumping a histogram of memory usage from a running JVM and see where >>> the memory is going. I can't think of anything in particular that >>> changed... >>> >>> >>> On 8/28/13 4:39 PM, Jeff Peters wrote: >>> >>>> >>>> I am tasked with updating our ancient (circa 7/10/2012) Giraph to >>>> giraph-release-1.0.0-RC3. Most jobs run fine but our largest job now runs >>>> out of memory using the same AWS elastic-mapreduce configuration we have >>>> always used. I have never tried to configure either Giraph or the AWS >>>> Hadoop. We build for Hadoop 1.0.2 because that's closest to the 1.0.3 AWS >>>> provides us. The 8 X m2.4xlarge cluster we use seems to provide 8*14=112 >>>> map tasks fitted out with 2GB heap each. Our code is completely unchanged >>>> except as required to adapt to the new Giraph APIs. Our vertex, edge, and >>>> message data are completely unchanged. On smaller jobs, that work, the >>>> aggregate heap usage high-water mark seems about the same as before, but >>>> the "committed heap" seems to run higher. I can't even make it work on a >>>> cluster of 12. In that case I get one map task that seems to end up with >>>> nearly twice as many messages as most of the others so it runs out of >>>> memory anyway. It only takes one to fail the job. Am I missing something >>>> here? Should I be configuring my new Giraph in some way I didn't used to >>>> need to with the old one? >>>> >>>> >>> >> >> >> > > >
