[ewg] Using ib_send_bw for multicast performance test

[Ido Shamai]

Hello Peter and sorry for the late reply.
My name is Ido and I'm the perftest maintainer.

As you suspect, the multicast feature in perftest is only intended for end 2 end machines, and he does not support multiple receivers vs. sender. But the changes should be minor in order to complete this task.
Please remove all ctx_hand_shake calls from the code, and supply as parameters the MGID manually (using -M) for both sides.
Instruct the sender to send to this MGID, and the server should automatically receive the packets.
Actually you can take only the multicast_resources.* code and use it to register MGID and then you will be able to construct a
much simple test for your need.

I can try code by myself for you (and maybe commit this to the package) but I cannot promise a time frame for completing this,
and as I say you should be able to do it yourself.

Make sure you are working with the latest perftest code (git://git.openfabrics.org/~shamoya/perftest.git).

I would be glad to help on any issue, please add the word perftest to the subject.

Regards,
Idos.


On 8/17/2011 1:10 PM, Kovacs Peter Tamas wrote:
> Hi all,
>
> I'm trying to benchmark multicast performance on a simple Infiniband network with three nodes using ib_send_bw, but apparently we cannot find the right parameters for doing so.
> Point-to-point measurements work fine, but when setting up a multicast test with one sender and multiple receivers, we cannot make them start. The best
> I guess it must be something like "ib_send_bw -c UD -g <some number> " on the sender side, and "ib_send_bw -c UD" on the receivers side, but I don't know how the senders and receivers can all connect to exchange data, like they do in a point-to-point test by providing the IP address of the listening end.
>
> Skimming through the source, ctx_hand_shake looks to me like it's for two parties, not more than two, but I must be wrong. The readme also talks about a single server and a single client. As far as I can see, I cannot list multiple IPs when using the multicast feature either. The best I could make is an "Failed to exchange date between server and clients" error message, making one of the listening servers exit.
> I've got predefined multicast groups from 0xc000 to 0xc0009, and queried the corresponding GIDs with saquery, but feeding any of these via the -M option did not help either.
>
> Could any of you please shed some light about using this feature, ie. measuring multicast performance between 1+N nodes, one sender and N receivers.
>
> With many thanks,
> Peter Kovacs
> _______________________________________________
> ewg mailing list
> ewg at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
>