Subject: Re: SS20 network performance
To: None <martin@duskware.de, port-sparc@netbsd.org>
From: None <eeh@netbsd.org>
List: port-sparc
Date: 05/17/2002 14:52:53
| > Now, the maximum throughput I've been able to achieve, is 2.57MB/s,
| > measured using wget
| ..
| > This is strange, because I've reached over 7 MB/s throughputs with
| > Solaris 8 running ipfilter on the very same platform.
|
| So this is interesting. 2.57MB/s is not bad, my LX only gets around
| 1.1MB/s.
|
| Matthias Scheeler once tried to address this by bringing in a version of
| the hme driver that uses pre-allocated dma maps, thereby avoiding one copy
| of the data.
|
| We then benchmarked this, and it didn't help performance on the LX (which may
| be due to the crippled cache size, of course) and it *dramatically* reduced
| performance on sparc64 machines (I get the expected 11MB/s on my U5 with hme).
|
| So this driver version was backed out again, until further research would
| have shown the culprit for the sparc64 performance issue (the iommu operations
| are suspect) and avoided it. AFAICT this has not happened yet.
|
| Other drivers using similar techniques (if_vr.c for example) exhibit the
| same performance problems on sparc64.

Have you considered kernel profiling to see what's causing the slowdown?

Eduardo