Re: PROBLEM: Alpha SMP Low Outbound Bandwidth

From: Urban Widmark (urban@teststation.com)
Date: Sat May 26 2001 - 06:17:47 EST


On 25 May 2001, Jay Thorne wrote:

> Netperf is a pretty good idea. Should not be a cpu bottleneck. Thats a
> good thing. So pretty much the same results as wu-ftpd: Note that I used
> the 466 mhz quad with a via-rhine, since the 400 locked up and was still
> fscking when I started this test.
>
> Recv Send Send
> Socket Socket Message Elapsed
> Size Size Size Time Throughput
> bytes bytes bytes secs. 10^6bits/sec
>
> To alpha 87380 16384 16384 10.02 39.25
> x86 local 87380 16384 16384 9.99 559.46
> alpha local 87380 16384 16384 10.01 547.27
> alp to x86 87380 16384 16384 10.01 25.77
> another x86 87380 16384 16384 9.99 553.67
> to same x86 87380 16384 16384 10.00 82.79
> and back 87380 16384 16384 10.00 93.89

What type of NIC is in the x86'es? If they are not the same, what happens
if you put one of those in the alphas?
(and what happens with the via-rhine if put in one of the x86'es?)

Alphas don't like unaligned memory accesses (not sure how bad those are).
I think that you can get some idea on the frequence of those with 'cat
/proc/cpuinfo' where it should say "kernel unaligned acc" and then some
numbers.

If that first number keeps going up then the driver(s) or something isn't
being nice. Well, just a though. Oh yes, your original message showed a
value of 1646246 for kernel space unaligned accesses. Is that high?

The userspace value is 0, so I assume that is high. The value for 'pc='
should give the address of where the last unaligned access took place
(look it up in System.map or /proc/ksyms)

Does the value grow faster when you run your netperf tests?

That still doesn't explain the SMP vs UP difference.

/Urban

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Thu May 31 2001 - 21:00:28 EST