Subject: Re: freebsd em driver
To: Frank van der Linden <firstname.lastname@example.org>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
Date: 06/17/2002 13:58:03
In message <20020617223753.A13260@vaasje.org>Frank van der Linden writes
>Ok, it might simply be that my setup (and that of a few others
>I talked to) isn't optimal, and that the numbers I am seeing
>are simply too low. It's been a while since I tested a wide
>variety of gig-e hardware.
>Btw, with a faster/wider PCI bus, I meant faster than
>32/33 which is still the default in most boards, although
>that's changing now.
I dont understand this at all. Every gigabit-ethernet board I've seen
has been 64-bit wide or 66mhz, or both, right back to the prototypes
for what became the original Wiseman. Trying to do performance tests
gigabit hardware is going to hit the limits of what the PCI bus and
(possibly) host memory, which has a guaranteed-not-to-exceed rate of
some 120Mbytes/sec; generally lower, in practice, what with PCI
arbitration and retries.
Last time I tried this on a 32/33 system, I hit about 640 Mbit/sec
throughput; as far as the console user could tell, we'd stalled the
Maybe I should try some numbers on a P4 desktop with 33/32 PCI; there,
the numbers you gave may be in the right ballpark. But for desktops
with 64/66 PCI (and memory/chipsets capable of reaching it), the
appropriate target is ca. 90 Mbytes/sec. (Higher, for the Broadom cards.)
>> That'd be interesting, esp. if I can collar enough machines to run
>> some head-to-head tests.
>Ok, I'll put up a tarfile.
Thanks, will look forward to it.
Will address the rest in private email.