Le 03/09/2010 15:38, George Bosilca a écrit :
> I think you will have to revert this patch as the btl_bandwidth __IS__ supposed to be in Mbs and not MBs. We usually talk about networks in Mbs (there is a pattern in Ethernet 1G/10G, Myricom 10G). In addition the original design of the multi-rail was based on this assumption, and the multi-rail handling code deal with these values (at that level I don't think it really matters, but at least it needs consistent values from all BTLs).
> However, going over the existing BTLs I can see that some BTLs do not correctly set this value:
> BTL Bandwidth Auto-detect Status
> Elan 2000 NO Correct
2000 looks strange to me. Last time I played with Elan4, bandwidth was
900MB/s or so.
> GM 250 NO Doubtful
> MX 2000/10000 YES (Mbs) Correct (before the patch)
> OFUD 800 NO Doubtful
> OpenIB 2000/4000/8000 YES (Mbs) Correct (multiplied by the active_width)
I found the problem when using both MX and OpenIB at the same time, so
they can't be both wrong or both correct. IB was reporting 800, not
2000/4000/8000. Maybe because auto-detect didn't work and the default is
btl_openib_mca.c:527: mca_btl_openib_module.super.btl_bandwidth = 800;