Re: Dual ported "E card" performance

Markus Fischer <[email protected]> Sun, 25 Jan 2004 16:53:19 -0800
Newsgroups gmane.network.myrinet.general
Message-ID <[email protected]>
Bill,

last year at SC2003 in Phoenix we demonstrated Sockets-GM performance using
2 E - Cards, on dual Xeon nodes. The combined sustained bandwith using 
two clients was
7.1 GBit/s.

Depending on the message size you can switch to a zero copy protocol. 
The reported
CPU usage is 5-7 per cent for large messages.

I will copy paste the netperf output which is using one E - Card:

[root@asa3 netperf]# ./netperf -l 5 -H 192.168.2.201 -- -m 4000 -M 4000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384   4000    5.00     3397.72

[root@asa3 netperf]# ./netperf -l 5 -H 192.168.2.201 -- -m 8000 -M 8000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send
Socket Socket  Message  Elapsed
Size   Size    Size     Time     Throughput
bytes  bytes   bytes    secs.    10^6bits/sec

 87380  16384   8000    5.00     3944.04


With CPU Usage:

[root@asa3 netperf]# ./netperf -c -C -l 5 -H 192.168.2.201 -- -m 125000 
-M 125000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send                          Utilization       Service 
Demand
Socket Socket  Message  Elapsed              Send     Recv     Send    Recv
Size   Size    Size     Time     Throughput  local    remote   local   
remote
bytes  bytes   bytes    secs.    10^6bits/s  % T      % T      us/KB   us/KB

 87380  16384 125000    5.00       3115.48   12.00    13.00    0.631   0.684

[root@asa3 netperf]# ./netperf -c -C -l 5 -H 192.168.2.201 -- -m 150000 
-M 150000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send                          Utilization       Service 
Demand
Socket Socket  Message  Elapsed              Send     Recv     Send    Recv
Size   Size    Size     Time     Throughput  local    remote   local   
remote
bytes  bytes   bytes    secs.    10^6bits/s  % T      % T      us/KB   us/KB

 87380  16384 150000    5.00       3226.62   11.40    10.20    0.579   0.518

[root@asa3 netperf]# ./netperf -c -C -l 5 -H 192.168.2.201 -- -m 250000 
-M 250000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send                          Utilization       Service 
Demand
Socket Socket  Message  Elapsed              Send     Recv     Send    Recv
Size   Size    Size     Time     Throughput  local    remote   local   
remote
bytes  bytes   bytes    secs.    10^6bits/s  % T      % T      us/KB   us/KB

 87380  16384 250000    5.00       3436.45   8.40     7.80     0.400   0.372

[root@asa3 netperf]# ./netperf -c -C -l 5 -H 192.168.2.201 -- -m 500000 
-M 500000
TCP STREAM TEST to 192.168.2.201
Recv   Send    Send                          Utilization       Service 
Demand
Socket Socket  Message  Elapsed              Send     Recv     Send    Recv
Size   Size    Size     Time     Throughput  local    remote   local   
remote
bytes  bytes   bytes    secs.    10^6bits/s  % T      % T      us/KB   us/KB

 87380  16384 500000    5.00       3592.06   5.80     7.80     0.265   0.356


Bill Hartner wrote:

>On a dual Xeon, the user sustained uni/bi-directional data rate for the
>dual ported "E card" is quoted as 495/950 MB/s.
>
>http://www.myri.com/news/03a14/
>
>This rate was for a GM 2.1 microbenchmark.  Are there perf results for a
>GM sockets application or netperf unmodified on the "E card" - including 
>throughput, cpu utilization, and latency ?  
>
>I am assuming you can assign one IP address to the dual ports.  Will
>the packets be distributed across both ports if TCP streaming to/from
>another dual card with a single ip address or TCP streaming to/from two
>other clients each with a single port "D" card using Sockets GM or just
>tcp/ip ?
>
>-Bill
>_______________________________________________
>Myrinet mailing list
>[email protected]
>http://email.osc.edu/mailman/listinfo/myrinet
>
>  
>