Here are my findings with TCP and MX. In fact for TCP results on
heterogeneous networks we should wait one or two days (still
upgrading my cluster). But I go some very interesting results for MX.
Thanks to Myricom guys for the access to their resources. I was able
to run Open MPI on a cluster with 2 Myricom interfaces, a 10Gbs and a
2Gbs. The green and violet lines are for the 2Gbs respectively for
the 10Gbs interfaces. The blue line is for the current version, as in
a fresh check-out. As you can see the bandwidth is getting a little
bit higher than the 10g, but still far from what it should be (in
numbers we get 9.3Gbs over the 10g, 1.9 over the 2g and 10Gb over
both of them). Then I start applying the patches. First, the Gleb
original patch, the one without the dynamic scheduling. This is the
red line. There seems to be no apparent difference between the trunk
and the version with this patch (there is barely few Mbs more with
the patch when I look at the numbers). Then I went on and added the
second patch, the one adding the dynamic routing. And there we got
some interesting results, i.e. the yellow line. We are at about 98%
of the theoretical bandwidth (i.e. the sum of the max over each
interfaces). Which is pretty good.
I think Gleb's second patch is the right way to go. Feel free to
commit as soon as you want.
On Jun 28, 2007, at 1:17 PM, Gleb Natapov wrote:
> On Thu, Jun 28, 2007 at 12:02:14PM -0400, George Bosilca wrote:
>> I'm not against the patch (at least not against your second version).
>> I really want to have the dynamic way to feed the BTLs based on the
>> order in which they complete the previous send. Give me one or two
>> days, I want to test your patch on a heterogeneous Ethernet
>> environment, and right now my cluster is enjoying an upgrade.
> No problem. I also will be glad if you'll be able to test with
> Myricom networks and mixed Myricom/ethernet environment.
>> On Jun 28, 2007, at 10:06 AM, Gleb Natapov wrote:
>>> Nobody except George haven't commented/complained about this patch,
>>> so I
>>> assume everybody except George are OK with it. And from George
>>> mails I
>>> don't understand if he is OK with me applying it to the trunk and
>>> he simply
>>> thinks that further work should be done in this area. So I'll ask
>>> directly: George are you OK with me putting the patch into the
>>> trunk? ;)
>>> devel mailing list
>> devel mailing list
> devel mailing list
- application/pkcs7-signature attachment: smime.p7s