[mvapich-discuss] Difference between MVAPICH2 and MVAPICH2-GDR
khaled hamidouche
hamidouc at cse.ohio-state.edu
Thu Jul 9 11:27:31 EDT 2015
Hi Makai,
Good to know that you have switched to MV2-GDR and be able to installed.
Regarding the performance behavior, it might be because the QPI bottleneck
as your GPU and HCA are in different socket. Can you please provide the
details about your node configuration ? and in meantime can you please try
to explicitly select the HCA and GPU on the same socket and try? you can
use the runtime parameter MV2_IBA_HCA.
Thanks a lot.
On Thu, Jul 9, 2015 at 10:55 AM, makai <makailove123 at 163.com> wrote:
> --===============6704199941958321163==
> Content-Type: multipart/alternative;
> boundary="Apple-Mail=_6426E929-5CE2-4F63-8FE2-EDF29E238E86"
>
> --Apple-Mail=_6426E929-5CE2-4F63-8FE2-EDF29E238E86
> Content-Transfer-Encoding: quoted-printable
> Content-Type: text/plain; charset="gb2312"
>
> I have installed MVAPICH2-GDR and gdrcopy, but when I run the =
> osu_latency, it turned out to be weird.
>
> makai at gpu-cluster-3:~$ $MV2_PATH/bin/mpiexec -hosts =
> 192.168.2.3,192.168.2.4 -n 2 -env MV2_USE_CUDA 1 =
> /opt/mvapich2/gdr/2.1/cuda6.5/gnu/libexec/mvapich2/osu_latency D D
> # OSU MPI-CUDA Latency Test
> # Send Buffer on DEVICE (D) and Receive Buffer on DEVICE (D)
> # Size Latency (us)
> Warning *** The GPU and IB selected are not on the same socket. Do not =
> delever the best performance=20
> Warning *** The GPU and IB selected are not on the same socket. Do not =
> delever the best performance=20
> 0 1.67
> 1 2.91
> 2 3.92
> 4 3.99
> 8 3.92
> 16 3.97
> 32 160.67
> 64 161.51
> 128 162.05
> 256 165.20
> 512 165.88
> 1024 168.92
> 2048 176.08
> 4096 185.95
> 8192 72.63
> 16384 261.26
> 32768 148.08
> 65536 518.37
> 131072 143.93
> 262144 260.03
> 524288 254.41
> 1048576 393.54
> 2097152 672.47
> 4194304 1244.69
>
> Why the result became so bad after size became larger than 32?
> And I find there is only one CA on my node, why it told me =A1=B0The GPU =
> and IB selected are not on the same socket. Do not delever the best =
> performance =A1=B0?
>
> Could you give me some help?
> Thanks!
>
> > =D4=DA 2015=C4=EA6=D4=C225=C8=D5=A3=AC=CF=C2=CE=E78:56=A3=ACPanda, =
> Dhabaleswar <panda at cse.ohio-state.edu> =D0=B4=B5=C0=A3=BA
> >=20
> > Thanks for your note. You are mixing-up two concepts: 1) CUDA-aware =
> MPI and 2) GPUDirect RDMA. The=20
> > CUDA-aware MPI concept allows MPI_Send and MPI_Recv to use data from =
> GPU device directly.=20
> > GPUDirect RDMA (GDR) allows data to be moved from a GPU to another GPU =
> through through PCI=20
> > interface using RDMA (say over InfiniBand) without going through the =
> host memory.=20
> >=20
> > MVAPICH2 supports only CUDA-aware MPI.
> >=20
> > MVAPICH2-GDR supports CUDA-aware MPI, GPUDirect RDMA (GDR), and many =
> other advanced designs
> > related to GPU clusters to exploit performance and scalability. For =
> example, you can get very low=20
> > D-D latency (close to 2 microsec) with MVAPICH2-GDR. Thus, for GPU =
> clusters with InfiniBand, =20
> > we strongly recommend the users to use MVAPICH2-GDR. Please take a =
> look at the MVAPICH2-GDR user=20
> > guide from the following URL for all features and usage guidelines:=20
> >=20
> > http://mvapich.cse.ohio-state.edu/userguide/gdr/
> >=20
> > Hope this helps.=20
> >=20
> > DK
> >=20
> >=20
> >=20
> > ________________________________________
> > From: mvapich-discuss-bounces at cse.ohio-state.edu on behalf of makai =
> [makailove123 at 163.com]
> > Sent: Thursday, June 25, 2015 1:02 AM
> > To: mvapich-discuss at cse.ohio-state.edu
> > Subject: [mvapich-discuss] Difference between MVAPICH2 and =
> MVAPICH2-GDR
> >=20
> > I have installed MVAPICH2, and it says that it supports GPUDirect =
> RDMA. MPI_Send and MPI_Recv could use addresses of device for data =
> transmission.
> > So, what=A1=AFs the difference between MVAPICH2 and MVAPICH2-GDR?
> >=20
> > _______________________________________________
> > mvapich-discuss mailing list
> > mvapich-discuss at cse.ohio-state.edu
> > http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
> > [attachment]
> >=20
> > winmail.dat
> > download: http://u.163.com/t0/r4YIVzg
> >=20
> > <winmail.dat>
>
>
> --Apple-Mail=_6426E929-5CE2-4F63-8FE2-EDF29E238E86
> Content-Transfer-Encoding: quoted-printable
> Content-Type: text/html; charset="gb2312"
>
> <html><head><meta http-equiv=3D"Content-Type" content=3D"text/html =
> charset=3Dgb2312"></head><body style=3D"word-wrap: break-word; =
> -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" =
> class=3D""><div class=3D"">I have installed MVAPICH2-GDR and gdrcopy, =
> but when I run the osu_latency, it turned out to be weird.</div><div =
> class=3D""><br class=3D""></div><div class=3D""><div style=3D"margin: =
> 0px; font-family: 'Andale Mono'; color: rgb(41, 249, 20); =
> background-color: rgb(0, 0, 0);" class=3D"">makai at gpu-cluster-3:~$ =
> $MV2_PATH/bin/mpiexec -hosts 192.168.2.3,192.168.2.4 -n 2 -env =
> MV2_USE_CUDA 1 =
> /opt/mvapich2/gdr/2.1/cuda6.5/gnu/libexec/mvapich2/osu_latency D =
> D</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; color: =
> rgb(41, 249, 20); background-color: rgb(0, 0, 0);" class=3D""># OSU =
> MPI-CUDA Latency Test</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D""># Send Buffer on DEVICE (D) and Receive Buffer on DEVICE =
> (D)</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; color: =
> rgb(41, 249, 20); background-color: rgb(0, 0, 0);" class=3D""># =
> Size Latency (us)</div><div =
> style=3D"margin: 0px; font-family: 'Andale Mono'; color: rgb(41, 249, =
> 20); background-color: rgb(0, 0, 0);" class=3D"">Warning *** The GPU and =
> IB selected are not on the same socket. Do not delever the best =
> performance </div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">Warning *** The GPU and IB selected are not on the same =
> socket. Do not delever the best performance </div><div =
> style=3D"margin: 0px; font-family: 'Andale Mono'; color: rgb(41, 249, =
> 20); background-color: rgb(0, 0, 0);" class=3D"">0 =
> =
> 1.67</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; color: =
> rgb(41, 249, 20); background-color: rgb(0, 0, 0);" class=3D"">1 =
> =
> 2.91</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; =
> color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" class=3D"">2 =
> =
> 3.92</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">4 =
> 3.99</div><div style=3D"margin: 0px; =
> font-family: 'Andale Mono'; color: rgb(41, 249, 20); background-color: =
> rgb(0, 0, 0);" class=3D"">8 =
> 3.92</div><div style=3D"margin: =
> 0px; font-family: 'Andale Mono'; color: rgb(41, 249, 20); =
> background-color: rgb(0, 0, 0);" class=3D"">16 =
> =
> 3.97</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; color: =
> rgb(41, 249, 20); background-color: rgb(0, 0, 0);" class=3D"">32 =
> =
> 160.67</div><div style=3D"margin: 0px; font-family: 'Andale Mono'; =
> color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">64 =
> 161.51</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">128 =
> 162.05</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">256 =
> 165.20</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">512 =
> 165.88</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">1024 =
> 168.92</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">2048 =
> 176.08</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">4096 =
> 185.95</div><div style=3D"margin: 0px; font-family: =
> 'Andale Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">8192 =
> 72.63</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">16384 =
> 261.26</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">32768 =
> 148.08</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">65536 =
> 518.37</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">131072 =
> 143.93</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">262144 =
> 260.03</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">524288 =
> 254.41</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">1048576 =
> 393.54</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">2097152 =
> 672.47</div><div style=3D"margin: 0px; font-family: 'Andale =
> Mono'; color: rgb(41, 249, 20); background-color: rgb(0, 0, 0);" =
> class=3D"">4194304 =
> 1244.69</div></div><div class=3D""><br class=3D""></div><div =
> class=3D"">Why the result became so bad after size became larger than =
> 32?</div><div class=3D""><span id=3D"x-apple-selection:end"></span>And I =
> find there is only one CA on my node, why it told me =A1=B0<span =
> style=3D"color: rgb(41, 249, 20); font-family: 'Andale Mono'; =
> background-color: rgb(0, 0, 0);" class=3D"">The GPU and IB selected are =
> not on the same socket. Do not delever the best performance</span><span =
> style=3D"color: rgb(41, 249, 20); font-family: 'Andale Mono'; =
> background-color: rgb(0, 0, 0);" class=3D""> </span>=A1=B0?</div><div=
> class=3D""><br class=3D""></div><div class=3D"">Could you give me some =
> help?</div><div class=3D"">Thanks!</div><br class=3D""><div><blockquote =
> type=3D"cite" class=3D""><div class=3D"">=D4=DA =
> 2015=C4=EA6=D4=C225=C8=D5=A3=AC=CF=C2=CE=E78:56=A3=ACPanda, Dhabaleswar =
> <<a href=3D"mailto:panda at cse.ohio-state.edu" =
> class=3D"">panda at cse.ohio-state.edu</a>> =D0=B4=B5=C0=A3=BA</div><br =
> class=3D"Apple-interchange-newline"><div class=3D"">Thanks for your =
> note. You are mixing-up two concepts: 1) CUDA-aware MPI and 2) GPUDirect =
> RDMA. The <br class=3D"">CUDA-aware MPI concept allows MPI_Send and =
> MPI_Recv to use data from GPU device directly. <br class=3D"">GPUDirect =
> RDMA (GDR) allows data to be moved from a GPU to another GPU through =
> through PCI <br class=3D"">interface using RDMA (say over InfiniBand) =
> without going through the host memory. <br class=3D""><br =
> class=3D"">MVAPICH2 supports only CUDA-aware MPI.<br class=3D""><br =
> class=3D"">MVAPICH2-GDR supports CUDA-aware MPI, GPUDirect RDMA (GDR), =
> and many other advanced designs<br class=3D"">related to GPU clusters to =
> exploit performance and scalability. For example, you can get very low =
> <br class=3D"">D-D latency (close to 2 microsec) with MVAPICH2-GDR. =
> Thus, for GPU clusters with InfiniBand, <br class=3D"">we strongly =
> recommend the users to use MVAPICH2-GDR. Please take a look at the =
> MVAPICH2-GDR user <br class=3D"">guide from the following URL for all =
> features and usage guidelines: <br class=3D""><br class=3D""><a =
> href=3D"http://mvapich.cse.ohio-state.edu/userguide/gdr/" =
> class=3D"">http://mvapich.cse.ohio-state.edu/userguide/gdr/</a><br =
> class=3D""><br class=3D"">Hope this helps. <br class=3D""><br =
> class=3D"">DK<br class=3D""><br class=3D""><br class=3D""><br =
> class=3D"">________________________________________<br class=3D"">From: =
> mvapich-discuss-bounces at cse.ohio-state.edu on behalf of makai =
> [makailove123 at 163.com]<br class=3D"">Sent: Thursday, June 25, 2015 1:02 =
> AM<br class=3D"">To: mvapich-discuss at cse.ohio-state.edu<br =
> class=3D"">Subject: [mvapich-discuss] Difference between MVAPICH2 and =
> MVAPICH2-GDR<br class=3D""><br class=3D"">I have installed MVAPICH2, and =
> it says that it supports GPUDirect RDMA. MPI_Send and MPI_Recv could use =
> addresses of device for data transmission.<br class=3D"">So, what=A1=AFs =
> the difference between MVAPICH2 and MVAPICH2-GDR?<br class=3D""><br =
> class=3D"">_______________________________________________<br =
> class=3D"">mvapich-discuss mailing list<br =
> class=3D"">mvapich-discuss at cse.ohio-state.edu<br =
> class=3D"">
> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-disc=
> uss<br class=3D"">[attachment]<br class=3D""><br class=3D"">winmail.dat<br=
> class=3D"">download: http://u.163.com/t0/r4YIVzg<br class=3D""><br =
> class=3D""><span =
> id=3D"cid:564DF121-75AF-47FA-986A-51CCA2B11238"><winmail.dat></span>=
> </div></blockquote></div><br class=3D""></body></html>=
>
> --Apple-Mail=_6426E929-5CE2-4F63-8FE2-EDF29E238E86--
>
> --===============6704199941958321163==
> Content-Type: text/plain; charset="us-ascii"
> MIME-Version: 1.0
> Content-Transfer-Encoding: 7bit
> Content-Disposition: inline
>
> _______________________________________________
> mvapich-discuss mailing list
> mvapich-discuss at cse.ohio-state.edu
> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>
> --===============6704199941958321163==--
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20150709/0d7585f4/attachment-0001.html>
More information about the mvapich-discuss
mailing list