Re: [Xen-users] XEN 4.3.1 network performance issue

["NiX" <nix@xxxxxxxxxxxxxxxx>]

> On Fri, Dec 27, 2013 at 02:17:19PM +0200, NiX wrote:
>> > On Tue, Dec 24, 2013 at 05:06:37PM +0200, NiX wrote:
>> >> Hi. I am running a XEN VM in PV mode. This VM has four cores enabled.
>> VM
>> >> network config is as follows:
>> >>
>> >> vif = ["ip=127.0.0.1,mac=XX:XX:XX:XX:XX:XX,bridge=br0,rate=100Mb/s"]
>> >>
>> >> Dom0 is dual XEON X5450 with 16GB of RAM.
>> >>
>> >> I can barely zmap at 10mbps ~ 12k packets/second on this VM
>> >>
>> >
>> > My google-fu tells me that zmap sends out gigantic amount of TCP SYN
>> > packets in its default configuration. It is possible the ring size
>> > between frontend and backend is the bottleneck. Data copying from
>> > frontend to backend might also be a problem.
>> >
>> > You can probably patch your Dom0 and DomU kernel with multi-page ring
>> > support to see if it makes it better.
>> >
>> > Wei.
>> >
>>
>> I was not able neither to exceed 10k packets/second with a lightweight
>> equivalent software and netback/0 fully utilized one 3GHz core.
>>
>
> Hmm... You're using kthread-based netback. In your use case frontend is
> likely to create interrupt storm for backend. Do you have breakdown of
> CPU usage (sy, si etc)?

I am using Linux nix 3.2.53-grsec #1 SMP and the server is pretty much
idle expect when I attempt to zmap.Load average: 0.09 0.06 0.13

ps aux | grep netback
root      1411  2.6  0.0      0     0 ?        S    Dec09 696:17 [netback/0]
root      1412  0.0  0.0      0     0 ?        S    Dec09   2:44 [netback/1]
root      1413  1.1  0.0      0     0 ?        S    Dec09 312:39 [netback/2]
root      1414  0.0  0.0      0     0 ?        S    Dec09   0:00 [netback/3]
root      1415  0.0  0.0      0     0 ?        S    Dec09   8:22 [netback/4]
root      1416  0.0  0.0      0     0 ?        S    Dec09   2:54 [netback/5]
root      1417  0.0  0.0      0     0 ?        S    Dec09   0:00 [netback/6]
root      1418  0.0  0.0      0     0 ?        S    Dec09   0:00 [netback/7]

It seems to only fully utilize netback/0 when I attempt to zmap and takes
one 3GHz core ... Maybe there's something to look at in terms of
optimizing?


>
>> Of course I had tweaked both host/vm network because the default limits
>> somewhat cannot properly handle 10k or more packets/second due to the
>> default 1k open files limit and so forth
>>
>> Well you know, basic users can manage with 1k packets and/or connections
>> second but I am not a basic user ;) Is there any handy link on how to
>> use
>> multi-page ring feature. I am going to test that.
>>
>
> Oops I missed the CPU load in your previous email. In that case CPU was
> already fully loaded so multi-page ring might not provide you much
> benefit. Plus there's no convenient way to test it so I think it doesn't
> worth your time. Sorry.
>
> Just courious, do you have any figures on baremetal performs for zmap?
>

Yes. I zmap using xeon X3220. The network is 1gbps and I use approx.
225Mbps of bandwidth ~ 350k packets/second

18:05 20% (1h12m left); send: 380081546 389 Kp/s (350 Kp/s avg); recv:
84362 38 p/s (77 p/s avg); drops: 32.0 Kp/s (20.9 Kp/s avg); hits: 0.02%

So at that rate, it takes about 1 hour and 45 minute to scan 2 billion IP's.
Please find attachment for CPU usage.


> Wei.
>

Attachment: zmap.png
Description: PNG image

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxx
http://lists.xen.org/xen-users