Parallelism doesn't really affect the throughput as long as it's: - not less than the number of available execution slots, - ... and probably some low multiple of them to even out task size effects - not so high that the bookkeeping overhead dominates
Although you may need to select different scales of parallelism for different stages (like a join), you shouldn't in general have to change it according to data size. However you could count the input size and make parallelism some function of that if you found that was consistently better. The one exception are operations that tend to pull data into memory. You may need more parallelism as scale increases to keep in-memory data size small enough. There again you usually just err on the side of 'too much' parallelism, or avoid patterns that can pull a lot of data into memory, but this is usually the pain point if there is one. The problem I run into when thinking about this is that I don't think Spark can do much better, since it doesn't have the info above needed to decide these things in general. The calling program has to tell it. On Wed, Mar 4, 2015 at 10:17 AM, Jeff Zhang <zjf...@gmail.com> wrote: > Hi Sean, > > > If you know a stage needs unusually high parallelism for example you can > repartition further for that stage. > > The problem is we may don't know whether high parallelism is needed. e.g. > for the join operator, high parallelism may only be necessary for some > dataset that lots of data can join together while for other dataset high > parallelism may not be necessary if only a few data can join together. > > So my question is that unable changing parallelism at runtime dynamically > may not be flexible. > > > > On Wed, Mar 4, 2015 at 5:36 PM, Sean Owen <so...@cloudera.com> wrote: >> >> Hm, what do you mean? You can control, to some extent, the number of >> partitions when you read the data, and can repartition if needed. >> >> You can set the default parallelism too so that it takes effect for most >> ops thay create an RDD. One # of partitions is usually about right for all >> work (2x or so the number of execution slots). >> >> If you know a stage needs unusually high parallelism for example you can >> repartition further for that stage. >> >> On Mar 4, 2015 1:50 AM, "Jeff Zhang" <zjf...@gmail.com> wrote: >>> >>> Thanks Sean. >>> >>> But if the partitions of RDD is determined before hand, it would not be >>> flexible to run the same program on the different dataset. Although for the >>> first stage the partitions can be determined by the input data set, for the >>> intermediate stage it is not possible. Users have to create policy to >>> repartition or coalesce based on the data set size. >>> >>> >>> On Tue, Mar 3, 2015 at 6:29 PM, Sean Owen <so...@cloudera.com> wrote: >>>> >>>> An RDD has a certain fixed number of partitions, yes. You can't change >>>> an RDD. You can repartition() or coalese() and RDD to make a new one >>>> with a different number of RDDs, possibly requiring a shuffle. >>>> >>>> On Tue, Mar 3, 2015 at 10:21 AM, Jeff Zhang <zjf...@gmail.com> wrote: >>>> > I mean is it possible to change the partition number at runtime. >>>> > Thanks >>>> > >>>> > >>>> > -- >>>> > Best Regards >>>> > >>>> > Jeff Zhang >>> >>> >>> >>> >>> -- >>> Best Regards >>> >>> Jeff Zhang > > > > > -- > Best Regards > > Jeff Zhang --------------------------------------------------------------------- To unsubscribe, e-mail: user-unsubscr...@spark.apache.org For additional commands, e-mail: user-h...@spark.apache.org
