Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] TIPC BTL code ready for review
From: teng ma (xiaok1981_at_[hidden])
Date: 2011-08-29 09:22:02


Is your interconnect Gigabytes Ethernet? It's very surprised to see TCP BTL
just got 33MBytes peak BW on your cluster. I did a similar test on an amd
cluster with gigabytes Ethernet. As following shows, the TCP BTL's BW is
similar with your tipc(112MBytes/s). Could you redo the test with 2
processes spawned, 2 nodes in your machinefile and enabling --bynode?

It looks like your tipc BTL is pretty good at message size between 8K and
512K. Can you tell us more about difference between TIPC and TCP protocol
stacks? Any special configure needed to enable your tipc? Maybe you can
write a module in Netpipe( similar to NPTcp )to test raw performance on both
TCP and TIPC without MPI.

TCP BTL on a Gigbytes ethernet
#---------------------------------------------------
# Benchmarking PingPong
# #processes = 2
#---------------------------------------------------
       #bytes #repetitions t[usec] Mbytes/sec
            0 1000 23.27 0.00
            1 1000 23.78 0.04
            2 1000 23.77 0.08
            4 1000 25.47 0.15
            8 1000 23.94 0.32
           16 1000 24.36 0.63
           32 1000 24.83 1.23
           64 1000 25.76 2.37
          128 1000 27.25 4.48
          256 1000 30.66 7.96
          512 1000 36.86 13.25
         1024 1000 49.00 19.93
         2048 1000 77.83 25.10
         4096 1000 82.42 47.39
         8192 1000 165.28 47.27
        16384 1000 325.01 48.08
        32768 1000 440.75 70.90
        65536 640 1060.00 58.96
       131072 320 1674.71 74.64
       262144 160 2814.13 88.84
       524288 80 4975.11 100.50
      1048576 40 9526.94 104.97
      2097152 20 18419.33 108.58
      4194304 10 36150.05 110.65
      8388608 5 71880.79 111.30

Teng

On Mon, Aug 29, 2011 at 3:51 AM, Xin He <xin.i.he_at_[hidden]> wrote:

>
>
>
> On 08/25/2011 03:14 PM, Jeff Squyres wrote:
>
>> On Aug 25, 2011, at 8:25 AM, Xin He wrote:
>>
>> Can you edit your configure.m4 directly and test it and whatnot? I
>>>> provided the configure.m4 as a starting point for you. :-) It shouldn't be
>>>> hard to make it check linux/tipc.h instead of tipc.h. I'm happy to give you
>>>> direct write access to the bitbucket, if you want it.
>>>>
>>> I think me having write access is convenient for both of us :)
>>>
>> Sure -- what's your bitbucket account ID?
>>
> It's "letter113"
>
> As we've discussed off-list, we can't take the code upstream until the
>>>> contributor agreement is signed, unfortunately.
>>>>
>>>> The agreement thing is ongoing right now, but it may take some time.
>>>
>> No worries. Lawyers tend to take time when reviewing this stuff; we've
>> seen this pattern in most organizations who sign the OMPI agreement.
>>
>> But to save time, can you guys do some test on TIPC BTL, so that
>>> when the agreement is ready, the code can be used?
>>>
>> I don't know if any of us has the TIPC support libraries installed.
>>
>
> It is easy to have TIPC support. It is within the kernel actually. To get
> TIPC working, you only have to configure it by using "tipc-config". Maybe
> you
> can check this doc for information: http://tipc.sourceforge.net/**
> doc/Users_Guide.txt <http://tipc.sourceforge.net/doc/Users_Guide.txt>
>
> So... what *is* TIPC? Is there a writeup anywhere that we can read about
>> what it is / how it works? For example, what makes TIPC perform better than
>> TCP?
>>
>
> Sure. Search "TIPC: Providing Communication for Linux Clusters". It is a
> paper written by the author of TIPC, explaining basic stuff about TIPC,
> should be very useful. And you can visit TIPC homepage:
> http://tipc.sourceforge.net/ .
>
> I have done some tests using tools like NetPIPE, osu and IMB and the
>>>>> result shows that TIPC BTL has a better performance
>>>>> than TCP BTL.
>>>>>
>>>> Great! Can you share any results?
>>>>
>>> Yes, please check the appendix for the results using IMB 3.2.
>>>
>>> I have done the tests on 2 computers. Dell SC1435
>>> Dual-Core AMD Opteron(tm) Processor 2212 HE x 2
>>> 4 GB Mem
>>> Ubuntu Server 10.04 LTS 32-bit Linux 2.6.32-24
>>>
>> I'm not familiar with the Dell or Opteron lines -- how recent are those
>> models?
>>
>> I ask because your TCP latency is a bit high (about 85us in 2-process IMB
>> PingPong); it might suggest older hardware. Or you may have built a
>> debugging version of Open MPI (if you have a .svn or .hg checkout, that's
>> the default). See the HACKING top-level file for how to get an optimized
>> build.
>>
>> For example, with my debug build of Open MPI on fairly old Xeons with 1GB
>> ethernet, I'm getting the following PingPong results (note: this is a debug
>> build; it's not even an optimized build):
>>
>> -----
>> $ mpirun --mca btl tcp,self --bynode -np 2 --mca btl_tcp_if_include eth0
>> hostname
>> svbu-mpi008
>> svbu-mpi009
>> $ mpirun --mca btl tcp,self --bynode -np 2 --mca btl_tcp_if_include eth0
>> IMB-MPI1 PingPong
>> #-----------------------------**----------------------
>> # Intel (R) MPI Benchmark Suite V3.2, MPI-1 part
>> #-----------------------------**----------------------
>> ...
>> #-----------------------------**----------------------
>> # Benchmarking PingPong
>> # #processes = 2
>> #-----------------------------**----------------------
>> #bytes #repetitions t[usec] Mbytes/sec
>> 0 1000 57.31 0.00
>> 1 1000 57.71 0.02
>> 2 1000 57.73 0.03
>> 4 1000 57.81 0.07
>> 8 1000 57.78 0.13
>> -----
>>
>> With an optimized build, it shaves off only a few us (which isn't too
>> important in this case, but it does matter in the low-latency transport
>> cases):
>>
>> -----
>> #-----------------------------**----------------------
>> # Benchmarking PingPong
>> # #processes = 2
>> #-----------------------------**----------------------
>> #bytes #repetitions t[usec] Mbytes/sec
>> 0 1000 54.62 0.00
>> 1 1000 54.92 0.02
>> 2 1000 55.15 0.03
>> 4 1000 55.16 0.07
>> 8 1000 55.15 0.14
>> -----
>>
>> Hi, I think these models are reasonably new :)
> The result I gave you, they are tested on 2 processes but on 2 different
> servers. I get that the result you showed is 2 processes on one machine?
> But I did build with debug enabled, I will try optimize then :)
>
> BTW, I forgot to tell you about SM & TIPC. Unfortunately, TIPC does not
> beat SM...
>
> /Xin
>
>
> ______________________________**_________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/**mailman/listinfo.cgi/devel>
>