[lkml]   [2002]   [May]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: Tcp/ip offload card driver
At 08:07 AM 10/05/2002 -0700, David S. Miller wrote:
>if there was some real life
>demonstrable performance gain with current systems.
>For example, do a SpecWEB run with TUX both using on-chip-TCP and
>without, same networking card. Show a demonstrable gain from the
>on-chip-TCP implementation. I bet you can't. If you can make such a
>claim using a setup that other people could reproduce themselves by
>buying your card and running the test, I'll eat all of my words.

i think its conceivable that TOE could be advantagous in some situations.

for example, if TOE was supported in a driver (eg. /dev/toeN) which allowed
user-space to mmap() into RAM (either on the card or main memory which the
TOE card DMAs into/out-of).
in that way, a read()-equvalent system-call on a TOE socket could
potentially just use 1 or 2 memory-accesses versus the 3 that currently occur.
translation: 33% or 50% less traffic across the front-side-bus, no
pollution of the processor L1/L2 cache as a side-effect of copy_to_user().

in terms of a write()-like system-call, the equivalent holds true also.
right now, a write() means 3 memory-accesses:
- 2 memory-accesses (memory-read and memory-write) on copy_from_user()
from userspace to kernel
- memory-read on NIC DMAing from kernel buffer
its forseeable that something that this could be 1 or 2 memory-accesses in
a kernel-memory-mmap()ed-to-userspace scenario.

the biggest advantage is the scenario where userspace doesn't care about
the actual payload of the data.
ie. its reading from-disk and sending to-network, or read-from-socket,

of course, if TOE adapters don't work this way and don't have some API of
their own exposed to allow this, then i'd agree that the benefit is negligible.
and of course, if we did get around to some kind of async i/o scheme in
linux that allowed the above, then TOE would be of limited value also.
of course, such a scheme would probably be too hard to play fair with the
page-cache, so it'd only be a "win" for data that is touched once and/or
the application needs to do its own page-cache-like function.

i guess this is some of the points i'm trying to show with high-speed Fibre
Channel HBAs -- the overhead of copy_to_user(), copy_from_user() means that
the front-side-bus/memory-bandwidth ultimately becomes the main bottleneck,
at least in IA32-architecture machines.



To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 13:22    [W:0.151 / U:0.404 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site