Hi,
I've just been comparing the performance of stunnel on Solaris 10 & Linux (Centos 5.5) and am seeing a big difference. I'm using stunnel 4.33 compiled with the same options on both platforms and the tcpstress benchmarking tool (http://stunnel.mirt.net/?page=perf).
On a Dell 780 running Solaris 10 I got 163.31 connections/second but on an older, slower Dell 620 running Linux I got 408.01 connections/second. Both machines were otherwise idle.
Does anyone know why there would be such a big difference between the two OS's?
Thanks,
Steve.
Steve Cliffe Senior Engineer - Development Infrastructure Team Andrew Network Solutions Asia Pacific
[ Address : Andrew Corporation Building (39), Northfields Avenue, Wollongong University, NSW, Australia 2500 ] [ Postal Address : PO Box U40, Wollongong University, NSW, Australia 2500 ] [ Tel : +61 2 4221 2991] [ Fax : +61 2 4221 2901 ] [ Email : [email protected]]
Visit Andrew on the Web at www.andrew.comhttp://www.andrew.com/
[cid:[email protected]][cid:[email protected]]
Cliffe, Stephen wrote:
Does anyone know why there would be such a big difference between the
two
OS's?
Linux kernel performed a huge rewrite and significant performance optimization of major subsystems (scheduler, networking) in 2.6 kernels. AFAIK Slowlaris was never really optimized for systems with less then 16 CPUs.
You may find this article interesting: http://www.ibm.com/developerworks/linux/library/l-web26/index.html
Mike
On Fri, Sep 17, 2010 at 02:49:25PM +0800, Cliffe, Stephen wrote:
Hi,
I've just been comparing the performance of stunnel on Solaris 10 & Linux (Centos 5.5) and am seeing a big difference. I'm using stunnel 4.33 compiled with the same options on both platforms and the tcpstress benchmarking tool (http://stunnel.mirt.net/?page=perf).
On a Dell 780 running Solaris 10 I got 163.31 connections/second but on an older, slower Dell 620 running Linux I got 408.01 connections/second. Both machines were otherwise idle.
How is your CPU usage? Can you post the output of the following commands:
% psrinfo | wc -l
% vmstat 1 10 (during the stress test)
% netstat -s (before and after the test)
CPU is around 95% idle during the test on Solaris, 35% on Linux.
Solaris Performance (Core 2 Duo 3.3GHz):
Benchmarking 10000 connection(s) by 1 thread(s) Bechmark done: success 66.02 seconds, 151.47 connections/second
Linux Performance (Pentium 4 CPU 3.4GHz dual-core):
Benchmarking 10000 connection(s) by 1 thread(s) Bechmark done: success 22.59 seconds, 442.72 connections/second
I've attached the vmstat and before & after netstat outputs.
Steve.
-----Original Message----- From: Jeremie Le Hen [mailto:[email protected]] Sent: Friday, 17 September 2010 7:46 PM To: Cliffe, Stephen Cc: [email protected] Subject: Re: [stunnel-users] Large performance difference: Solaris vs Linux
On Fri, Sep 17, 2010 at 02:49:25PM +0800, Cliffe, Stephen wrote:
Hi,
I've just been comparing the performance of stunnel on Solaris 10 & Linux (Centos 5.5) and am seeing a big difference. I'm using stunnel 4.33 compiled with the same options on both platforms and the tcpstress benchmarking tool (http://stunnel.mirt.net/?page=perf).
On a Dell 780 running Solaris 10 I got 163.31 connections/second but on an older, slower Dell 620 running Linux I got 408.01 connections/second. Both machines were otherwise idle.
How is your CPU usage? Can you post the output of the following commands:
% psrinfo | wc -l
% vmstat 1 10 (during the stress test)
% netstat -s (before and after the test)
Given that both boxes are single CPU, albeit dual-core, and there are 4 processes running for the benchmark (tcpcli, tcpserv, 2 * stunnel) I figured a single thread test would be reasonable.
I did repeat the benchmark with 10 threads and the slower Linux box still outperformed the faster Solaris box, although not by as much.
Steve.
-----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Michal Trojnara Sent: Monday, 20 September 2010 6:40 PM To: [email protected] Subject: Re: [stunnel-users] Large performance difference: Solaris vs Linux
Cliffe, Stephen wrote:
Benchmarking 10000 connection(s) by 1 thread(s)
Are you sure benchmarking with one thread (= single connection at a time) reflects your environment?
Mike _______________________________________________ stunnel-users mailing list [email protected] http://stunnel.mirt.net/mailman/listinfo/stunnel-users
-
Cliffe, Stephen -
Jeremie Le Hen -
Michal Trojnara