Hello guys, This is a really interesting discussion. I am also trying to fine-tune the performance of my cluster and especially my end-to-end-latency which ranges from 200-1200 msec for a topology with 2 spouts (each one with 2k tuples per second input rate) and 3 bolts. My cluster consists of 3 zookeeper nodes (1 shared with nimbus) and 6 supervisor nodes, all of them being AWS m4.xlarge instances.
I am pretty sure that the latency I am experiencing is ridiculous and I currently have no ideas what to do to improve that. I have 3 workers per node, which I will drop it to one worker per node after this discussion and see if I have better results. Thanks, Nick On Mon, Oct 5, 2015 at 10:40 AM, Kashyap Mhaisekar <[email protected]> wrote: > Anshu, > My methodology was as follows. Since the true parallelism of a machine is > the the no. of cores, I set the workers equal to no. of cores. (5 in my > case). That being said, since we have 32 GB per box, we usually leave 50% > off leaving us 16 GB spread across 5 machines. Hence we set the worker heap > at 3g. > > This was before Javiers and Michaels suggestion of keeping one JVM per > node... > > Ours is a single topology running on the boxes and hence I would be > changing it to one JVM (worker) per box and rerunning. > > Thanks > Kashyap > > On Mon, Oct 5, 2015 at 9:18 AM, anshu shukla <[email protected]> > wrote: > >> Sorry for reposting !! Any suggestions Please . >> >> Just one query How we can map - >> *1-no of workers to number of cores * >> *2-no of slots on one machine to number of cores over that machine* >> >> On Mon, Oct 5, 2015 at 7:32 PM, John Yost <[email protected]> >> wrote: >> >>> Hi Javier, >>> >>> Gotcha, I am seeing the same thing, and I see a ton of worker restarts >>> as well. >>> >>> Thanks >>> >>> --John >>> >>> On Mon, Oct 5, 2015 at 9:01 AM, Javier Gonzalez <[email protected]> >>> wrote: >>> >>>> I don't have numbers, but I did see a very noticeable degradation of >>>> throughput and latency when using multiple workers per node with the same >>>> topology. >>>> On Oct 5, 2015 7:25 AM, "John Yost" <[email protected]> wrote: >>>> >>>>> Hi Everyone, >>>>> >>>>> I am curious--are there any benchmark numbers that demonstrate how >>>>> much better one worker per node is? The reason I ask is that I may need >>>>> to >>>>> double up the workers on my cluster and I was wondering how much of a >>>>> throughput hit I may take from having two workers per node. >>>>> >>>>> Any info would be very much appreciated--thanks! :) >>>>> >>>>> --John >>>>> >>>>> >>>>> >>>>> On Sat, Oct 3, 2015 at 9:04 AM, Javier Gonzalez <[email protected]> >>>>> wrote: >>>>> >>>>>> I would suggest sticking with a single worker per machine. It makes >>>>>> memory allocation easier and it makes inter-component communication much >>>>>> more efficient. Configure the executors with your parallelism hints to >>>>>> take >>>>>> advantage of all your availabe CPU cores. >>>>>> >>>>>> Regards, >>>>>> JG >>>>>> >>>>>> On Sat, Oct 3, 2015 at 12:10 AM, Kashyap Mhaisekar < >>>>>> [email protected]> wrote: >>>>>> >>>>>>> Hi, >>>>>>> I was trying to come up with an approach to evaluate the parallelism >>>>>>> needed for a topology. >>>>>>> >>>>>>> Assuming I have 5 machines with 8 cores and 32 gb. And my topology >>>>>>> has one spout and 5 bolts. >>>>>>> >>>>>>> 1. Define one worker port per CPU to start off. (= 8 workers per >>>>>>> machine ie 40 workers over all) >>>>>>> 2. Each worker spawns one executor per component per worker, it >>>>>>> translates to 6 executors per worker which is 40x6= 240 executors. >>>>>>> 3. Of this, if the bolt logic is CPU intensive, then leave >>>>>>> parallelism hint at 40 (total workers), else increase parallelism hint >>>>>>> beyond 40 till you hit a number beyond which there is no more visible >>>>>>> performance. >>>>>>> >>>>>>> Does this look right? >>>>>>> >>>>>>> Thanks >>>>>>> Kashyap >>>>>>> >>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Javier González Nicolini >>>>>> >>>>> >>>>> >>> >> >> >> -- >> Thanks & Regards, >> Anshu Shukla >> > > -- Nikolaos Romanos Katsipoulakis, University of Pittsburgh, PhD student
