httperf throughput

httperf throughput

[Apparao, Padmashree K]

[-- Attachment #1.1: Type: text/plain, Size: 693 bytes --]

Hi,

 

I am running apache and httperf on the latest Xen unstable.

 

When I have the webserver running in a single cpu guest I can barely get
720 Mbps over 2 nics (each 1 Gb). That is the throughput is about 36% of
the native linux throughput. (I get line rate on native linux with the
cpu being fully utilized). The guest cpu is 100% utilized while dom0 is
about 80% utilized.

 

Said another way, I can get 2000 connections per second in Xen before
the cpu chokes while in Linux I get about 7000 connections per sec. Is
this what is expected. 

 

Does anyone have similar/alternate performance data to share?

Thanks

-          Padma

-            

 


[-- Attachment #1.2: Type: text/html, Size: 4884 bytes --]

[-- Attachment #2: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: httperf throughput

[Nivedita Singhvi]

Apparao, Padmashree K wrote:
> Hi,

> I am running apache and httperf on the latest Xen unstable.

> When I have the webserver running in a single cpu guest I can barely get
> 720 Mbps over 2 nics (each 1 Gb). That is the throughput is about 36% of
> the native linux throughput. (I get line rate on native linux with the
> cpu being fully utilized). The guest cpu is 100% utilized while dom0 is
> about 80% utilized.
> 
>  
> 
> Said another way, I can get 2000 connections per second in Xen before
> the cpu chokes while in Linux I get about 7000 connections per sec. Is
> this what is expected. 
> 
>  
> 
> Does anyone have similar/alternate performance data to share?
> 
> Thanks
> 
> -          Padma

Padma,

The degradation from native for Xen guests will vary considerably
with workload.  There have not been too many network-intensive
workloads run yet, possibly because that's the least optimized
so far, although of late a lot of Herbert Xu's work has made
quite a difference.

How many procs in the box, and are you pinning the vcpus?
What are the parameters you are running with for httpperf?

I'd still recommend netperf over ttcp, which some of the other
people running tests are using, but I don't have any numbers.

Could you profile your benchmark runs and share the data? Would
be very interesting. Also, a little more information on the
system (memory, etc).


thanks,
Nivedita