[mpich-discuss] Scalability of Intel quad core (Harpertown) cluster

[Scott Atchley]

On Mar 28, 2008, at 1:50 AM, Hee Il Kim wrote:

> - If the gigabit network was the cause, could it be improved with  
> Open-MX?

Hi Kim,

Depending on your Ethernet switch, Open-MX should be able to lower  
your latency and reduce your CPU usage. I see about 10.5 us  
internally using the e1000 driver with interrupt coalescing turned  
off, others report that they have seen as low as 6.5 us (I do not  
know what NICs they used or if it the machines were switched or back- 
to-back).

To use Open-MX, your switch must support flow control for send and  
receive (some don't) and jumbo frames. If you need interoperability  
with native MX over Ethernet (some hosts with Myri-10G NICs connected  
to an Ethernet switch, some hosts with non-Myricom NICs), then your  
switch must handle at least 4 KB frames. If you do not need  
interoperability, then you can use 8 or 9 KB frames and possibly see  
slightly better performance (fewer interrupts).

Brice released version 0.4.0 which has working interoperability with  
native MX and is nearly feature complete. See http://www.open-mx.org  
for the status and source code.

Scott