Batch size impacts convergence, so bigger batch means more iterations. There are some approaches to deal with it (such as http://www.cs.cmu.edu/~muli/file/minibatch_sgd.pdf), but they need to be implemented and tested.
Nonetheless, could you share your thoughts regarding reducing this overhead in Spark (or probably a workaround)? Sorry for repeating it, but I think this is crucial for MLlib in Spark, because Spark is intended for bigger amounts of data. Machine learning with bigger data usually requires SGD (vs batch GD), SGD requires a lot of updates, and “Spark overhead” times “many updates” equals impractical time needed for learning. From: Shivaram Venkataraman [mailto:[email protected]] Sent: Sunday, April 05, 2015 7:13 PM To: Ulanov, Alexander Cc: [email protected]; Joseph Bradley; [email protected] Subject: Re: Stochastic gradient descent performance Yeah, a simple way to estimate the time for an iterative algorithms is number of iterations required * time per iteration. The time per iteration will depend on the batch size, computation required and the fixed overheads I mentioned before. The number of iterations of course depends on the convergence rate for the problem being solved. Thanks Shivaram On Thu, Apr 2, 2015 at 2:19 PM, Ulanov, Alexander <[email protected]<mailto:[email protected]>> wrote: Hi Shivaram, It sounds really interesting! With this time we can estimate if it worth considering to run an iterative algorithm on Spark. For example, for SGD on Imagenet (450K samples) we will spend 450K*50ms=62.5 hours to traverse all data by one example not considering the data loading, computation and update times. One may need to traverse all data a number of times to converge. Let’s say this number is equal to the batch size. So, we remain with 62.5 hours overhead. Is it reasonable? Best regards, Alexander From: Shivaram Venkataraman [mailto:[email protected]<mailto:[email protected]>] Sent: Thursday, April 02, 2015 1:26 PM To: Joseph Bradley Cc: Ulanov, Alexander; [email protected]<mailto:[email protected]> Subject: Re: Stochastic gradient descent performance I haven't looked closely at the sampling issues, but regarding the aggregation latency, there are fixed overheads (in local and distributed mode) with the way aggregation is done in Spark. Launching a stage of tasks, fetching outputs from the previous stage etc. all have overhead, so I would say its not efficient / recommended to run stages where computation is less than 500ms or so. You could increase your batch size based on this and hopefully that will help. Regarding reducing these overheads by an order of magnitude it is a challenging problem given the architecture in Spark -- I have some ideas for this, but they are very much at a research stage. Thanks Shivaram On Thu, Apr 2, 2015 at 12:00 PM, Joseph Bradley <[email protected]<mailto:[email protected]>> wrote: When you say "It seems that instead of sample it is better to shuffle data and then access it sequentially by mini-batches," are you sure that holds true for a big dataset in a cluster? As far as implementing it, I haven't looked carefully at GapSamplingIterator (in RandomSampler.scala) myself, but that looks like it could be modified to be deterministic. Hopefully someone else can comment on aggregation in local mode. I'm not sure how much effort has gone into optimizing for local mode. Joseph On Thu, Apr 2, 2015 at 11:33 AM, Ulanov, Alexander <[email protected]<mailto:[email protected]>> wrote: > Hi Joseph, > > > > Thank you for suggestion! > > It seems that instead of sample it is better to shuffle data and then > access it sequentially by mini-batches. Could you suggest how to implement > it? > > > > With regards to aggregate (reduce), I am wondering why it works so slow in > local mode? Could you elaborate on this? I do understand that in cluster > mode the network speed will kick in and then one can blame it. > > > > Best regards, Alexander > > > > *From:* Joseph Bradley > [mailto:[email protected]<mailto:[email protected]>] > *Sent:* Thursday, April 02, 2015 10:51 AM > *To:* Ulanov, Alexander > *Cc:* [email protected]<mailto:[email protected]> > *Subject:* Re: Stochastic gradient descent performance > > > > It looks like SPARK-3250 was applied to the sample() which GradientDescent > uses, and that should kick in for your minibatchFraction <= 0.4. Based on > your numbers, aggregation seems like the main issue, though I hesitate to > optimize aggregation based on local tests for data sizes that small. > > > > The first thing I'd check for is unnecessary object creation, and to > profile in a cluster or larger data setting. > > > > On Wed, Apr 1, 2015 at 10:09 AM, Ulanov, Alexander < > [email protected]<mailto:[email protected]>> wrote: > > Sorry for bothering you again, but I think that it is an important issue > for applicability of SGD in Spark MLlib. Could Spark developers please > comment on it. > > > -----Original Message----- > From: Ulanov, Alexander > Sent: Monday, March 30, 2015 5:00 PM > To: [email protected]<mailto:[email protected]> > Subject: Stochastic gradient descent performance > > Hi, > > It seems to me that there is an overhead in "runMiniBatchSGD" function of > MLlib's "GradientDescent". In particular, "sample" and "treeAggregate" > might take time that is order of magnitude greater than the actual gradient > computation. In particular, for mnist dataset of 60K instances, minibatch > size = 0.001 (i.e. 60 samples) it take 0.15 s to sample and 0.3 to > aggregate in local mode with 1 data partition on Core i5 processor. The > actual gradient computation takes 0.002 s. I searched through Spark Jira > and found that there was recently an update for more efficient sampling > (SPARK-3250) that is already included in Spark codebase. Is there a way to > reduce the sampling time and local treeRedeuce by order of magnitude? > > Best regards, Alexander > > --------------------------------------------------------------------- > To unsubscribe, e-mail: > [email protected]<mailto:[email protected]> > For additional commands, e-mail: > [email protected]<mailto:[email protected]> > > >
