Hello Bence, So, you have 96 cores at your disposal. My guess would be that 3 workers are not enough to use all of them, you should either try with a lot more, or try to multithread them as Avery said (thus, try 4 workers with 24 threads each). However, as I already reported, I tried this myself and I didn't notice any improvement in performance. I guess scaling with the number of worker threads is fundamental for a BSP framework so that's really weird, I guess I must be doing something wrong.
Could you please try increasing your workers and tell me if you also don't notice any improvement in performance? Furthermore, have you tried running with both 1 and 3 workers? Did you see any difference there? I really want to sort this scalability issue out... On a final note, please don't use the word "node" for describing a lot of different things, it got me quite confused :-) Your cluster's *nodes* have * cores* and your input is a graph with 140k *vertices. *Cheers, Alexandros On 29 November 2012 14:02, Magyar, Bence (US SSA) < bence.mag...@baesystems.com> wrote: > Folks, **** > > ** ** > > I have some of the same questions as Alexandros below. What is exactly is > “a worker”? I am not sure I understood Avery’s answer below. I have > 4-node cluster. Each node has 24 nodes. My first node is functioning (in > MapReduce parlance) as both a “job tracker” as well as a “task tracker”. > So I have 4 compute nodes. (I have verified that master/slave config is > correct). I am launching the Giraph SimpleShortestPathsVertex example on > an input graph with approximately 140,000 nodes/ 410,000 edges and the > computation is taking approx. 6 minutes. Although I don’t know what a > “good” number is, 6 minutes seems rather “slow” given all the compute > horsepower I have at my disposal. When I monitor “top” on my machines > while the compute is running, my cores are ~ 80-90% idle.**** > > ** ** > > I am launching my job with the following parameters:**** > > ** ** > > ./giraph -Dgiraph.useSuperstepCounters=false > -DSimpleShortestPathsVertex.sourceId=100 ../target/giraph.jar > org.apache.giraph.examples.SimpleShortestPathsVertex -if > org.apache.giraph.io.JsonLongDoubleFloatDoubleVertexInputFormat -ip > /user/hduser/in -of > org.apache.giraph.io.JsonLongDoubleFloatDoubleVertexOutputFormat -op > /user/hduser/out -w 3**** > > ** ** > > Note that I have my number of workers (–w =3). Should this be some other > value? Does anyone have any simple configuration suggestions that will > help me tune Giraph to my problem?**** > > ** ** > > Thanks! **** > > ** ** > > Bence**** > > ** ** > > *From:* Alexandros Daglis [mailto:alexandros.dag...@epfl.ch] > *Sent:* Thursday, November 29, 2012 6:19 AM > *To:* user@giraph.apache.org > *Subject:* Re: What a "worker" really is and other interesting runtime > information**** > > ** ** > > Ok, so I added the partitions flag, going with > > hadoop jar target/giraph-0.1-jar-with-dependencies.jar > org.apache.giraph.examples.SimpleShortestPathsVertex > -Dgiraph.SplitMasterWorker=false -Dgiraph.numComputeThreads=12 > -Dhash.userPartitionCount=12 input output 12 1 > > but still I got no overall speedup at all (compared to using 1 thread) and > only 1 out of 12 cores is utilized at most times. Isn't Giraph supposed to > exploit parallelism to get some speedup? Any other suggestion? > > Thanks, > Alexandros**** > > On 29 November 2012 00:20, Avery Ching <ach...@apache.org> wrote:**** > > Oh, forgot one thing. You need to set the number of partitions to use > single each thread works on a single partition at a time. > > Try -Dhash.userPartitionCount=<number of threads>**** > > > > On 11/28/12 5:29 AM, Alexandros Daglis wrote:**** > > Dear Avery, > > I followed your advice, but the application seems to be totally > thread-count-insensitive: I literally observe zero scaling of performance, > while I increase the thread count. Maybe you can point out if I am doing > something wrong. > > - Using only 4 cores on a single node at the moment > - Input graph: 14 million vertices, file size is 470 MB > - Running SSSP as follows: hadoop jar > target/giraph-0.1-jar-with-dependencies.jar > org.apache.giraph.examples.SimpleShortestPathsVertex > -Dgiraph.SplitMasterWorker=false -Dgiraph.numComputeThreads=X input output > 12 1 > where X=1,2,3,12,30 > - I notice a total insensitivity to the number of thread I specify. > Aggregate core utilization is always approximately the same (usually around > 25-30% => only one of the cores running) and overall execution time is > always the same (~8 mins) > > Why is Giraph's performance not scaling? Is the input size / number of > workers inappropriate? It's not an IO issue either, because even during > really low core utilization, time is wasted on idle, not on IO. > > Cheers, > Alexandros > > > **** > > On 28 November 2012 11:13, Alexandros Daglis <alexandros.dag...@epfl.ch> > wrote:**** > > Thank you Avery, that helped a lot! > > Regards, > Alexandros **** > > ** ** > > On 27 November 2012 20:57, Avery Ching <ach...@apache.org> wrote:**** > > Hi Alexandros, > > The extra task is for the master process (a coordination task). In your > case, since you are using a single machine, you can use a single task. > > -Dgiraph.SplitMasterWorker=false > > and you can try multithreading instead of multiple workers. > > -Dgiraph.numComputeThreads=12 > > The reason why cpu usage increases is due to netty threads to handle > network requests. By using multithreading instead, you should bypass this. > > Avery **** > > > > On 11/27/12 9:40 AM, Alexandros Daglis wrote:**** > > Hello everybody, > > I went through most of the documentation I could find for Giraph and also > most of the messages in this email list, but still I have not figured out > precisely what a "worker" really is. I would really appreciate it if you > could help me understand how the framework works. > > At first I thought that a worker has a one-to-one correspondence to a map > task. Apparently this is not exactly the case, since I have noticed that if > I ask for x workers, the job finishes after having used x+1 map tasks. What > is this extra task for? > > I have been trying out the example SSSP application on a single node with > 12 cores. Giving an input graph of ~400MB and using 1 worker, around 10 GBs > of memory are used during execution. What intrigues me is that if I use 2 > workers for the same input (and without limiting memory per map task), > double the memory will be used. Furthermore, there will be no improvement > in performance. I rather notice a slowdown. Are these observations normal? > > Might it be the case that 1 and 2 workers are very few and I should go to > the 30-100 range that is the proposed number of mappers for a conventional > MapReduce job? > > Finally, a last observation. Even though I use only 1 worker, I see that > there are significant periods during execution where up to 90% of the 12 > cores computing power is consumed, that is, almost 10 cores are used in > parallel. Does each worker spawn multiple threads and dynamically balances > the load to utilize the available hardware? > > Thanks a lot in advance! > > Best, > Alexandros > > **** > > ** ** > > ** ** > > ** ** > > ** ** > > ** ** >
