netdev
[Top] [All Lists]

Re: zerocopy results on GigE

To: blah@xxxxxxxxx (Benjamin C.R. LaHaise)
Subject: Re: zerocopy results on GigE
From: kuznet@xxxxxxxxxxxxx
Date: Thu, 8 Feb 2001 21:39:43 +0300 (MSK)
Cc: davem@xxxxxxxxxx, netdev@xxxxxxxxxxx
In-reply-to: <Pine.LNX.3.96.1010207161320.13604B-100000@kanga.kvack.org> from "Benjamin C.R. LaHaise" at Feb 7, 1 04:18:11 pm
Sender: owner-netdev@xxxxxxxxxxx
Hello!

> How many fragments are there in the typical tx packet?

2 or 3 for mtu=1500 (66% of two fragment and 33% of three fragment)

Well, you may count this assuming that header (ethernet +
ip + tcp/udp) is the first fragment, and payload consists of 4K pages.


>                                                 Are these machines
> with 64 bit / 66 MHz PCI, or just plain old 32 bit 33 MHz?  If it's a 33
> MHz bus, the overhead for starting new PCI cycles to gather fragments
> could be maxing out the bandwidth.  A PCI analyzer could be helpful =)

Yes, all the experimental data are about 32/33 pci.

Actually, it is not pci latency, but plain straight acenic bug.
F.e. e1000 does not have this overhead and reachs pps rates
at least twice more than acenic not straining muscles. 8)

>=5 usecs per dma transaction is something wicked,
but it is not a fatal flaw. At least, it should not be sensed
by TCP, it is parallel to software/protocol latencies and completely
hidden.

Alexey

<Prev in Thread] Current Thread [Next in Thread>