netdev
[Top] [All Lists]

Re: [ipv4, e1000] multi client throughput testing

To: "David S. Miller" <davem@xxxxxxxxxxxxx>
Subject: Re: [ipv4, e1000] multi client throughput testing
From: Jesse Brandeburg <jesse.brandeburg@xxxxxxxxx>
Date: Thu, 16 Jun 2005 17:48:23 -0700
Cc: netdev@xxxxxxxxxxxxxxx, shemminger@xxxxxxxx, jheffner@xxxxxxx, netdev@xxxxxxxxxxx
In-reply-to: <20050610.171127.59653238.davem@xxxxxxxxxxxxx>
Organization: Intel Corporation
References: <Pine.WNT.4.63.0506101507450.2756@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> <20050610.171127.59653238.davem@xxxxxxxxxxxxx>
Sender: netdev-bounce@xxxxxxxxxxx
User-agent: Mozilla Thunderbird 1.0.2 (Windows/20050317)
Ick, I get to be the bearer of my own bad news. I seem to mostly have a client misconfiguration problem.

David S. Miller wrote:
> From: Jesse Brandeburg <jesse.brandeburg@xxxxxxxxx>
> Date: Fri, 10 Jun 2005 16:56:50 -0700 (Pacific Daylight Time)
>
>  > What did i miss?
>
> Thanks for all of the data Jesse.  I'll try to sift through it this
> weekend.

Well, as it turns out I was sort of right all along, when i was thinking that the client's tcp windows were not being serviced quickly enough. First, I figured out that the windows client machines have a good "out of the box" behavior when receiving tcp data from linux. Second, the clients sending data to the server were maxing out their tcp window at 64k and did *not* have rfc1323 enabled. After enabling rfc1323 and upping the max window size to 128k, each client's throughput went up quite a bit (there may be more headroom i didn't test yet). Total throughput for us in this case is around 1560Mb/s now. I'd like to see it at 1700-1800 but I don't think it will do it. We're still running almost entirely in interrupt mode (with NAPI enabled) at about 7-8000 ints/s

Now I will go back and run with the netfilter enabled kernel and take a look again at the faster replenish/fairness patches I've been working on.

Thanks for your attention,
 Jesse

<Prev in Thread] Current Thread [Next in Thread>
  • Re: [ipv4, e1000] multi client throughput testing, Jesse Brandeburg <=