May 2020 - discuss - mailman3.cels.anl.gov

Better alternative to MPI_Allreduce() and avoiding deadlock with MPI_Neighbor_alltoallw().
by hritikesh semwal 05 May '20

05 May '20

On Tue, 5 May, 2020, 10:30 PM , <discuss-request(a)mpich.org> wrote: > Send discuss mailing list submissions to > discuss(a)mpich.org > > To subscribe or unsubscribe via the World Wide Web, visit > https://lists.mpich.org/mailman/listinfo/discuss > or, via email, send a message with subject or body 'help' to > discuss-request(a)mpich.org > > You can reach the person managing the list at > discuss-owner(a)mpich.org > > When replying, please edit your Subject line so it is more specific > than "Re: Contents of discuss digest..." > > > Today's Topics: > > 1. Better alternatives of MPI_Allreduce() (Benson Muite) > > > ---------------------------------------------------------------------- > > Message: 1 > Date: Tue, 05 May 2020 15:17:08 +0300 > From: "Benson Muite" <benson_muite(a)emailplus.org> > To: "Benson Muite via discuss" <discuss(a)mpich.org> > Subject: [mpich-discuss] Better alternatives of MPI_Allreduce() > Message-ID: <ef00483f-25a7-48d8-a04c-964e8001def7(a)www.fastmail.com> > Content-Type: text/plain; charset="us-ascii" > > > 1. I am using MPI_Neighbor_alltoallw() for exchanging the data by > generating a distributed graph topology communicator. My concern is that > most of the time my code is working fine but sometimes I guess it is going > into deadlock (as it is not showing any output). But MPI_Neighbor_alltoallw > uses MPI_Waitall inside it so I am not getting why exactly this is > happening. > >> > >> May want to check sending and receiving correct data. Perhaps also try > MPI_Neighbor_alltoallw > >> > >> > 2. Is it possible that every time I run the code the processors > times for completion of the task may vary? For example, for one run it all > processors takes around 100 seconds and for another run, all processors > take 110 seconds. > >> > >> There is usually some variability. Do you solve the same system each > time? What is the method of solution? If your code is available it can > sometimes be easier to give suggestions. > >> > > Yes, the system of equations are the same. I am using the finite volume > method for solving Navier stokes equations. By first sentence you mean to > say it is possible. > > Is the method implicit or explicit? > Its an explicit method. > > > > >> > > >> > Please help in above two matters. > >> > > >> > On Tue, May 5, 2020 at 4:28 PM hritikesh semwal < > hritikesh.semwal(a)gmail.com> wrote: > >> >> Thanks for your response. > >> >> > >> >> Yes, you are right. I have put barrier just before Allreduce and > out of the total time consumed by Allreduce, 79% time is consumed by the > barrier. But my computational work is balanced. Right now, I have > distributed 97336 cells among 24 processors and maximum and minimum cell > distribution among all processors is 4057 and 4055 respectively which is > not too bad. Is there any solution to get rid of this. > >> > >> Try profiling your code not just looking at cell distribution. Are any > profling tools already installed on your cluster? > > > > gprof and valgrind are there. > > While not ideal GPROF may be helpful. Perhaps initial try running on 12 > processors. With GPROF you will get 12 files to examine. Check if all > subroutines take similar times on each processor. You can also time the > subroutines individually using MPI_WTIME to get the same information. > Yes, I have already timed my code before posting this question. I will try with gprof. > > Also, try not to reply to the digest -, or if you do, change the subject > of the message. This is useful in deciding what to read. > Is it fine this time? I have changed the subject line. Is that what you want to say?

1 0

Better alternatives of MPI_Allreduce()
by Benson Muite 05 May '20

05 May '20

> 1. I am using MPI_Neighbor_alltoallw() for exchanging the data by generating a distributed graph topology communicator. My concern is that most of the time my code is working fine but sometimes I guess it is going into deadlock (as it is not showing any output). But MPI_Neighbor_alltoallw uses MPI_Waitall inside it so I am not getting why exactly this is happening. >> >> May want to check sending and receiving correct data. Perhaps also try MPI_Neighbor_alltoallw >> >> > 2. Is it possible that every time I run the code the processors times for completion of the task may vary? For example, for one run it all processors takes around 100 seconds and for another run, all processors take 110 seconds. >> >> There is usually some variability. Do you solve the same system each time? What is the method of solution? If your code is available it can sometimes be easier to give suggestions. >> > Yes, the system of equations are the same. I am using the finite volume method for solving Navier stokes equations. By first sentence you mean to say it is possible. Is the method implicit or explicit? > >> > >> > Please help in above two matters. >> > >> > On Tue, May 5, 2020 at 4:28 PM hritikesh semwal <hritikesh.semwal(a)gmail.com> wrote: >> >> Thanks for your response. >> >> >> >> Yes, you are right. I have put barrier just before Allreduce and out of the total time consumed by Allreduce, 79% time is consumed by the barrier. But my computational work is balanced. Right now, I have distributed 97336 cells among 24 processors and maximum and minimum cell distribution among all processors is 4057 and 4055 respectively which is not too bad. Is there any solution to get rid of this. >> >> Try profiling your code not just looking at cell distribution. Are any profling tools already installed on your cluster? > > gprof and valgrind are there. While not ideal GPROF may be helpful. Perhaps initial try running on 12 processors. With GPROF you will get 12 files to examine. Check if all subroutines take similar times on each processor. You can also time the subroutines individually using MPI_WTIME to get the same information. Also, try not to reply to the digest -, or if you do, change the subject of the message. This is useful in deciding what to read.

1 0

Re: [mpich-discuss] discuss Digest, Vol 91, Issue 5
by hritikesh semwal 05 May '20

05 May '20

On Tue, May 5, 2020 at 5:17 PM <discuss-request(a)mpich.org> wrote: > Send discuss mailing list submissions to > discuss(a)mpich.org > > To subscribe or unsubscribe via the World Wide Web, visit > https://lists.mpich.org/mailman/listinfo/discuss > or, via email, send a message with subject or body 'help' to > discuss-request(a)mpich.org > > You can reach the person managing the list at > discuss-owner(a)mpich.org > > When replying, please edit your Subject line so it is more specific > than "Re: Contents of discuss digest..." > > > Today's Topics: > > 1. Better alternatives of MPI_Allreduce() (Benson Muite) > 2. Re: Better alternatives of MPI_Allreduce() (Benson Muite) > 3. Re: Better alternatives of MPI_Allreduce() (hritikesh semwal) > > > ---------------------------------------------------------------------- > > Message: 1 > Date: Tue, 05 May 2020 14:20:57 +0300 > From: "Benson Muite" <benson_muite(a)emailplus.org> > To: "Benson Muite via discuss" <discuss(a)mpich.org> > Cc: hritikesh.semwal(a)gmail.com > Subject: [mpich-discuss] Better alternatives of MPI_Allreduce() > Message-ID: <a38d489f-25a6-40a1-9640-42bd2f52a80f(a)www.fastmail.com> > Content-Type: text/plain; charset="us-ascii" > > > >> > > >> > Hi Hitesh, > >> > > >> > What hardware are you running on and what is the interconnect? > > > > Right now I am using a cluster. > > What is the interconnect? > > > >> > Have you tried changing any of the MPI settings? > > > > What do you mean by MPI settings? > Given your comment on the barrier, this is probably not so useful at the > moment. > > > >> > Can the reduction be done asynchronously? > > > > I did not get your question. > > For example using a non blocking all reduce: > https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report/node135.htm > > > > >> > > >> > Regards, > >> > Benson > >> > >> Also, is your work load balanced? One way to check this might be to > place a barrier just before the all-reduce call. If the barrier ends up > taking most of your time, then it is likely you will need to determine a > better way to distribute the computational work. > > > > Thanks for your response. > > > > Yes, you are right. I have put barrier just before Allreduce and out of > the total time consumed by Allreduce, 79% time is consumed by the barrier. > But my computational work is balanced. Right now, I have distributed 97336 > cells among 24 processors and maximum and minimum cell distribution among > all processors is 4057 and 4055 respectively which is not too bad. Is there > any solution to get rid of this. >

1 0

Re: [mpich-discuss] Better alternatives of MPI_Allreduce()
by Benson Muite 05 May '20

05 May '20

On Tue, May 5, 2020, at 2:46 PM, hritikesh semwal wrote: > > > On Tue, May 5, 2020 at 4:51 PM Benson Muite <benson_muite(a)emailplus.org> wrote: >> __ >> >>>> > >>>> > Hi Hitesh, >>>> > >>>> > What hardware are you running on and what is the interconnect? >>> >>> Right now I am using a cluster. >> >> What is the interconnect? > > I don't know about this. Is it relevant? It can affect performance - but expect it may not be the most important factor on 24 processors. Most common is Infiniband (https://en.wikipedia.org/wiki/InfiniBand) > >>> >>>> > Have you tried changing any of the MPI settings? >>> >>> What do you mean by MPI settings? >> Given your comment on the barrier, this is probably not so useful at the moment. >>> >>>> > Can the reduction be done asynchronously? >>> >>> I did not get your question. >> >> For example using a non blocking all reduce: >> https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report/node135.htm >> > > I tried using a non-blocking call but after this code is not working correctly. Ok. Change back to blocking call. It is likely you have poor load balancing. >> >>> >>>> > >>>> > Regards, >>>> > Benson >>>> >>>> Also, is your work load balanced? One way to check this might be to place a barrier just before the all-reduce call. If the barrier ends up taking most of your time, then it is likely you will need to determine a better way to distribute the computational work. >>> >>> Thanks for your response. >>> >>> Yes, you are right. I have put barrier just before Allreduce and out of the total time consumed by Allreduce, 79% time is consumed by the barrier. But my computational work is balanced. Right now, I have distributed 97336 cells among 24 processors and maximum and minimum cell distribution among all processors is 4057 and 4055 respectively which is not too bad. Is there any solution to get rid of this? > Please help me in this regard. If you cannot profile your code, time the section before the all reduce on each processor using MPI_WTIME and check if it is even across all 24 processors. If using more processors, you will likely want to use a profiling tool, but if expect to run on about 24 processors, setting up a profiling tool if not already available may take some time.

1 0

Re: [mpich-discuss] Better alternatives of MPI_Allreduce()
by hritikesh semwal 05 May '20

05 May '20

On Tue, May 5, 2020 at 4:51 PM Benson Muite <benson_muite(a)emailplus.org> wrote: > > > > > Hi Hitesh, > > > > What hardware are you running on and what is the interconnect? > > > Right now I am using a cluster. > > > What is the interconnect? > I don't know about this. Is it relevant? > > > > Have you tried changing any of the MPI settings? > > > What do you mean by MPI settings? > > Given your comment on the barrier, this is probably not so useful at the > moment. > > > > > Can the reduction be done asynchronously? > > > I did not get your question. > > > For example using a non blocking all reduce: > https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report/node135.htm > > I tried using a non-blocking call but after this code is not working correctly. > > > > > > Regards, > > Benson > > Also, is your work load balanced? One way to check this might be to place > a barrier just before the all-reduce call. If the barrier ends up taking > most of your time, then it is likely you will need to determine a better > way to distribute the computational work. > > > Thanks for your response. > > Yes, you are right. I have put barrier just before Allreduce and out of > the total time consumed by Allreduce, 79% time is consumed by the barrier. > But my computational work is balanced. Right now, I have distributed 97336 > cells among 24 processors and maximum and minimum cell distribution among > all processors is 4057 and 4055 respectively which is not too bad. Is there > any solution to get rid of this? > > Please help me in this regard. > > > >

1 0

Re: [mpich-discuss] Better alternatives of MPI_Allreduce()
by Benson Muite 05 May '20

05 May '20

On Tue, May 5, 2020, at 2:16 PM, hritikesh semwal via discuss wrote: > I want to add two more questions about my solver, > 1. I am using MPI_Neighbor_alltoallw() for exchanging the data by generating a distributed graph topology communicator. My concern is that most of the time my code is working fine but sometimes I guess it is going into deadlock (as it is not showing any output). But MPI_Neighbor_alltoallw uses MPI_Waitall inside it so I am not getting why exactly this is happening. May want to check sending and receiving correct data. Perhaps also try MPI_Neighbor_alltoallw > 2. Is it possible that every time I run the code the processors times for completion of the task may vary? For example, for one run it all processors takes around 100 seconds and for another run, all processors take 110 seconds. There is usually some variability. Do you solve the same system each time? What is the method of solution? If your code is available it can sometimes be easier to give suggestions. > > Please help in above two matters. > > On Tue, May 5, 2020 at 4:28 PM hritikesh semwal <hritikesh.semwal(a)gmail.com> wrote: >> Thanks for your response. >> >> Yes, you are right. I have put barrier just before Allreduce and out of the total time consumed by Allreduce, 79% time is consumed by the barrier. But my computational work is balanced. Right now, I have distributed 97336 cells among 24 processors and maximum and minimum cell distribution among all processors is 4057 and 4055 respectively which is not too bad. Is there any solution to get rid of this. Try profiling your code not just looking at cell distribution. Are any profling tools already installed on your cluster? >> On Tue, May 5, 2020 at 12:30 PM Joachim Protze <protze(a)itc.rwth-aachen.de> wrote: >>> Hello, >>> >>> it is important to understand, that most of the time you see is not the >>> cost of the allreduce, but the cost of synchronization (caused by load >>> imbalance). >>> >>> You can do an easy experiment and add a barrier before the allreduce. >>> Then you will see the actual cost of the allreduce, while the cost of >>> synchronization will go into the barrier. >>> >>> Now, think about dependencies in your algorithm: do you need the output >>> value immediately? Is this the same time, where you have the input value >>> ready? >>> -> otherwise use non-blocking communication and perform independent work >>> in between >>> >>> In any case: fix your load imbalance (the root cause of synchronization >>> cost). >>> >>> Best >>> Joachim >>> >>> Am 05.05.20 um 07:38 schrieb hritikesh semwal via discuss: >>> > Hello all, >>> > >>> > I am working on the development of a parallel CFD solver and I am using >>> > MPI_Allreduce for the global summation of the local errors calculated on >>> > all processes of a group and the summation is to be used by all the >>> > processes. My concern is that MPI_Allreduce is taking almost 27-30% of >>> > the total time used, which is a significant amount. So, I want to ask if >>> > anyone can suggest me better alternative/s to replace MPI_Allreduce >>> > which can reduce the time consumption. >>> > >>> > Thank you. >>> > >>> > _______________________________________________ >>> > discuss mailing list discuss(a)mpich.org >>> > To manage subscription options or unsubscribe: >>> > https://lists.mpich.org/mailman/listinfo/discuss >>> > >>> >>> >>> -- >>> Dipl.-Inf. Joachim Protze >>> >>> IT Center >>> Group: High Performance Computing >>> Division: Computational Science and Engineering >>> RWTH Aachen University >>> Seffenter Weg 23 >>> D 52074 Aachen (Germany) >>> Tel: +49 241 80- 24765 >>> Fax: +49 241 80-624765 >>> protze(a)itc.rwth-aachen.de >>> www.itc.rwth-aachen.de >>> > _______________________________________________ > discuss mailing list discuss(a)mpich.org > To manage subscription options or unsubscribe: > https://lists.mpich.org/mailman/listinfo/discuss >

1 0

Better alternatives of MPI_Allreduce()
by Benson Muite 05 May '20

05 May '20

>> > >> > Hi Hitesh, >> > >> > What hardware are you running on and what is the interconnect? > > Right now I am using a cluster. What is the interconnect? > >> > Have you tried changing any of the MPI settings? > > What do you mean by MPI settings? Given your comment on the barrier, this is probably not so useful at the moment. > >> > Can the reduction be done asynchronously? > > I did not get your question. For example using a non blocking all reduce: https://www.mpi-forum.org/docs/mpi-3.1/mpi31-report/node135.htm > >> > >> > Regards, >> > Benson >> >> Also, is your work load balanced? One way to check this might be to place a barrier just before the all-reduce call. If the barrier ends up taking most of your time, then it is likely you will need to determine a better way to distribute the computational work. > > Thanks for your response. > > Yes, you are right. I have put barrier just before Allreduce and out of the total time consumed by Allreduce, 79% time is consumed by the barrier. But my computational work is balanced. Right now, I have distributed 97336 cells among 24 processors and maximum and minimum cell distribution among all processors is 4057 and 4055 respectively which is not too bad. Is there any solution to get rid of this.

1 0

Re: [mpich-discuss] Better alternatives of MPI_Allreduce()
by hritikesh semwal 05 May '20

05 May '20

I want to add two more questions about my solver, 1. I am using MPI_Neighbor_alltoallw() for exchanging the data by generating a distributed graph topology communicator. My concern is that most of the time my code is working fine but sometimes I guess it is going into deadlock (as it is not showing any output). But MPI_Neighbor_alltoallw uses MPI_Waitall inside it so I am not getting why exactly this is happening. 2. Is it possible that every time I run the code the processors times for completion of the task may vary? For example, for one run it all processors takes around 100 seconds and for another run, all processors take 110 seconds. Please help in above two matters. On Tue, May 5, 2020 at 4:28 PM hritikesh semwal <hritikesh.semwal(a)gmail.com> wrote: > Thanks for your response. > > Yes, you are right. I have put barrier just before Allreduce and out of > the total time consumed by Allreduce, 79% time is consumed by the barrier. > But my computational work is balanced. Right now, I have distributed 97336 > cells among 24 processors and maximum and minimum cell distribution among > all processors is 4057 and 4055 respectively which is not too bad. Is there > any solution to get rid of this. > > On Tue, May 5, 2020 at 12:30 PM Joachim Protze <protze(a)itc.rwth-aachen.de> > wrote: > >> Hello, >> >> it is important to understand, that most of the time you see is not the >> cost of the allreduce, but the cost of synchronization (caused by load >> imbalance). >> >> You can do an easy experiment and add a barrier before the allreduce. >> Then you will see the actual cost of the allreduce, while the cost of >> synchronization will go into the barrier. >> >> Now, think about dependencies in your algorithm: do you need the output >> value immediately? Is this the same time, where you have the input value >> ready? >> -> otherwise use non-blocking communication and perform independent work >> in between >> >> In any case: fix your load imbalance (the root cause of synchronization >> cost). >> >> Best >> Joachim >> >> Am 05.05.20 um 07:38 schrieb hritikesh semwal via discuss: >> > Hello all, >> > >> > I am working on the development of a parallel CFD solver and I am using >> > MPI_Allreduce for the global summation of the local errors calculated >> on >> > all processes of a group and the summation is to be used by all the >> > processes. My concern is that MPI_Allreduce is taking almost 27-30% of >> > the total time used, which is a significant amount. So, I want to ask >> if >> > anyone can suggest me better alternative/s to replace MPI_Allreduce >> > which can reduce the time consumption. >> > >> > Thank you. >> > >> > _______________________________________________ >> > discuss mailing list discuss(a)mpich.org >> > To manage subscription options or unsubscribe: >> > https://lists.mpich.org/mailman/listinfo/discuss >> > >> >> >> -- >> Dipl.-Inf. Joachim Protze >> >> IT Center >> Group: High Performance Computing >> Division: Computational Science and Engineering >> RWTH Aachen University >> Seffenter Weg 23 >> D 52074 Aachen (Germany) >> Tel: +49 241 80- 24765 >> Fax: +49 241 80-624765 >> protze(a)itc.rwth-aachen.de >> www.itc.rwth-aachen.de >> >>

1 0

Re: [mpich-discuss] discuss Digest, Vol 91, Issue 3
by hritikesh semwal 05 May '20

05 May '20

On Tue, May 5, 2020 at 12:30 PM <discuss-request(a)mpich.org> wrote: > Send discuss mailing list submissions to > discuss(a)mpich.org > > To subscribe or unsubscribe via the World Wide Web, visit > https://lists.mpich.org/mailman/listinfo/discuss > or, via email, send a message with subject or body 'help' to > discuss-request(a)mpich.org > > You can reach the person managing the list at > discuss-owner(a)mpich.org > > When replying, please edit your Subject line so it is more specific > than "Re: Contents of discuss digest..." > > > Today's Topics: > > 1. Better alternatives of MPI_Allreduce() (hritikesh semwal) > 2. Re: Better alternatives of MPI_Allreduce() (Benson Muite) > 3. Re: Better alternatives of MPI_Allreduce() (Benson Muite) > 4. Re: Better alternatives of MPI_Allreduce() (Joachim Protze) > > > ---------------------------------------------------------------------- > > Message: 1 > Date: Tue, 5 May 2020 11:08:20 +0530 > From: hritikesh semwal <hritikesh.semwal(a)gmail.com> > To: discuss(a)mpich.org > Subject: [mpich-discuss] Better alternatives of MPI_Allreduce() > Message-ID: > <CAA+35d1zgxVjcfH= > jpio5rLm1FkjxMC3Bis-qB9D43qjfEogRw(a)mail.gmail.com> > Content-Type: text/plain; charset="utf-8" > > Hello all, > > I am working on the development of a parallel CFD solver and I am using > MPI_Allreduce for the global summation of the local errors calculated on > all processes of a group and the summation is to be used by all the > processes. My concern is that MPI_Allreduce is taking almost 27-30% of the > total time used, which is a significant amount. So, I want to ask if anyone > can suggest me better alternative/s to replace MPI_Allreduce which can > reduce the time consumption. > > Thank you. >

1 0

Re: [mpich-discuss] Better alternatives of MPI_Allreduce()
by hritikesh semwal 05 May '20

05 May '20

Thanks for your response. Yes, you are right. I have put barrier just before Allreduce and out of the total time consumed by Allreduce, 79% time is consumed by the barrier. But my computational work is balanced. Right now, I have distributed 97336 cells among 24 processors and maximum and minimum cell distribution among all processors is 4057 and 4055 respectively which is not too bad. Is there any solution to get rid of this. On Tue, May 5, 2020 at 12:30 PM Joachim Protze <protze(a)itc.rwth-aachen.de> wrote: > Hello, > > it is important to understand, that most of the time you see is not the > cost of the allreduce, but the cost of synchronization (caused by load > imbalance). > > You can do an easy experiment and add a barrier before the allreduce. > Then you will see the actual cost of the allreduce, while the cost of > synchronization will go into the barrier. > > Now, think about dependencies in your algorithm: do you need the output > value immediately? Is this the same time, where you have the input value > ready? > -> otherwise use non-blocking communication and perform independent work > in between > > In any case: fix your load imbalance (the root cause of synchronization > cost). > > Best > Joachim > > Am 05.05.20 um 07:38 schrieb hritikesh semwal via discuss: > > Hello all, > > > > I am working on the development of a parallel CFD solver and I am using > > MPI_Allreduce for the global summation of the local errors calculated on > > all processes of a group and the summation is to be used by all the > > processes. My concern is that MPI_Allreduce is taking almost 27-30% of > > the total time used, which is a significant amount. So, I want to ask if > > anyone can suggest me better alternative/s to replace MPI_Allreduce > > which can reduce the time consumption. > > > > Thank you. > > > > _______________________________________________ > > discuss mailing list discuss(a)mpich.org > > To manage subscription options or unsubscribe: > > https://lists.mpich.org/mailman/listinfo/discuss > > > > > -- > Dipl.-Inf. Joachim Protze > > IT Center > Group: High Performance Computing > Division: Computational Science and Engineering > RWTH Aachen University > Seffenter Weg 23 > D 52074 Aachen (Germany) > Tel: +49 241 80- 24765 > Fax: +49 241 80-624765 > protze(a)itc.rwth-aachen.de > www.itc.rwth-aachen.de > >

1 0