We use haproxy and EC2 instances as load balancers for our clusters. The
tuning we use is pretty standard (somaxconn, nf_conntrack_max,
tcp_fin_timeout, rmem_max, wmem_max and etc) running vanilla ubuntu
AMIs. While EC2's instances and network have performance problems it is
possible to get reasonable reliability and performance from them. We
push between 10s of Mbps through a single c1.medium without issues, not
sure about beyond that.
-Joe
On 1/31/10 3:14 PM, Willy Tarreau wrote:
Hi Alexander,
On Sun, Jan 31, 2010 at 11:36:02PM +0100, Alexander Staubo wrote:
Has anyone any experience tuning HAProxy for performance when running
on Amazon EC2 instances? For example, are there any kernel parameters
that should be tuned differently, or are some instance types better
than others? Does HAProxy generally perform well on EC2?
well, last year I helped some guys in charge of a world wide sports
event which was hosted there. The performance was terrible. Completely
unstable. It was impossible to tune anything. Ping times would vary a
lot. It was impossible to know where the bottlenecks were, because
every machine was showing limited performance in turn without
necessarily having its CPU saturated. It was noticed that the internal
network was at least faulty, because the observed network congestions
were not constat and moving between machines. Sometimes it was even
almost impossible to type in SSH. We also discovered that when they
bought new nodes, some of them were under massive attacks, most likely
because people who are attacked quickly drop the nodes with the IPs
that belong to them and create new ones. So the attacked ones will
be picked by the next customer... Finally they moved to a real hosting
company with real machines and real performance in order to be able
to participate at least to a little part of the event.
In this experience, I think that for them, everything was virtual :
the machines, the network, the support, the availability, the visitors
and finally the profit.
I really can't say what you could play on to improve quality. After
having spent 3 full nights working with them on their machine, no
sensible trend appeared whatever we did. I think the real knobs are
outside your scope, on the other side of the VM :-/
Regards,
Willy
--
Name: Joseph A. Williams
Email: [email protected]
Blog: http://www.joeandmotorboat.com/
