Re: HAproxy performance on virtual sys. KVM

Hi Konrad,

On Fri, Jul 09, 2010 at 12:07:05PM +0200, Konrad wrote:
> Hello,
>
> I have a question about HAproxy configuration. I have 2 physical servers:
> 1) ServerA: Debian Lenny 64bit Kernel : 2.6.26-2-amd64 CPU: AMD
> Athlon(tm) 64 X2 RAM: 2GB
> 2) ServerB: Debian Lenny 64bit Kernel: 2.6.26-2-amd64 CPU: AMD
> Athlon(tm) 64 X2 RAM: 4GB
>
> On ServerA and ServerB I got 4 vm's:
> 1) ServerA: vm1 (HAproxy+keepalived) vm11 (installed: Apache - standard
> conf.) - MEM: 700MB Same OS as above
> 2) ServerB: vm2 (HAproxy+keepalived) vm21 (installed: Apache - standard
> conf.) - MEM: 700MB Same OS as above
>
> The network configuration for vm's is bridged - br0:
>
> haproxy:
> global
> log 127.0.0.1 local0
> log 127.0.0.1 local1 notice
> maxconn 4096
> user haproxy
> group haproxy
> daemon
> defaults
> log global
> mode http
> option httplog
> option dontlognull
> retries 3
> option redispatch
> maxconn 2000
> contimeout 5000
> clitimeout 50000
> srvtimeout 50000
>
> listen webfarm 192.168.0.111:80
> mode http
> stats enable
> stats auth admin:111
> balance roundrobin
> cookie JSESSIONID prefix
> option abortonclose
> option httpclose
> option forwardfor
> option httpchk HEAD /check.txt HTTP/1.0
> server webA 192.168.0.52:80 cookie A maxconn 512 check
> server webB 192.168.0.62:80 cookie B maxconn 512 check
>
>
> sysctl.conf (vm1, vm2)
> net.ipv4.tcp_max_syn_backlog = 10240
> net.ipv4.tcp_max_tw_buckets = 400000
> net.ipv4.ip_nonlocal_bind=1
> net.ipv4.tcp_fin_timeout=30
> net.ipv4.tcp_tw_reuse=1
> net.ipv4.tcp_max_orphans = 60000
> net.ipv4.tcp_synack_retries = 3
> net.core.somaxconn = 10000
>
> Test:
> a) vm11 (ab -c 100 -n 10000 http://127.0.0.1/check.txt):
> About: Requests per second: 5286.85 [#/sec] (mean)
> b) from 3 server (ab -c 100 -n 10000 http://192.168.0.111/check.txt):
> About: Requests per second: 986.80 [#/sec] (mean)
>
> I googled a lot and tuned the HAproxy configuration as above. Also the
> system configs but the performance didn't increase (~ 1k to 5k quite poor).
> I can send also any additional configurations if it's needed.

It would be worth checking from the host OS to the VMs too. Also, are you running iptables or anything stateful on the hosts ? I've never tried KVM yet, but even my (supposedly low-end) lguest performs a lot better on my netbook. Reason why I suppose something is blocking between your VMs.

Also, with "ab", you could also inject on haproxy's stats page, it will simply return you 401 and you'll see if the req/s sensibly increases or not. That way you'll know if the bottleneck is between ab and haproxy, between haproxy and the servers, or balanced between the two sides.

Regards,
Willy Received on 2010/07/10 10:11