Re: New benchmark of HAProxy at 10 Gbps using Myricom's 10GbE NICs possible?

From: Willy Tarreau <>
Date: Thu, 8 Sep 2011 07:51:42 +0200

Hi Aleks,

On Thu, Sep 08, 2011 at 12:52:20AM +0200, Aleksandar Lazic wrote:
> Hi Willy,
> I have take a look about the last test on
> and thought it would be nice to see a test with the brand new 1.5
> version.
> What do you think ;-)?

I agree in principle, but you know, running benchmarks takes a *lot* of time. The main issue I'm having right now is that I don't have a 10Gb switch, so I'm running with direct-attached machines, with only one client, one proxy and one server. In such a setup, it's common to see all 3 machines at 100% CPU during some tests, and it's hard to tell the part caused by client, server or proxy. The worst case I had was to be limited to about 15 Gbps full duplex and not being able to determine which machine or NIC was causing this, as most PCI-e chipsets I've encountered in the past were not able to reach the line rate.

I really need to build a lab with at least 5 machines (2 clients, 2 servers, one proxy and a switch). This is the only way to measure the impact on each component.

Anyway, what I'm expecting from new 10G runs with 1.5 is to have the same data rate over the wire (since we're doing nothing for this, we just use splice() to ask the kernel to do the work for us), but see an increase in the amount of peak connections per second. This is especially important for systems dealing with DDoSes using stick-tables, as they need to get rid of a connection as soon as possible.

Willy Received on 2011/09/08 07:51

This archive was generated by hypermail 2.2.0 : 2011/09/08 08:00 CEST