Re: [Xen-users] Xen version 4.3.2-rc1 netback CPU usage is too high

["H. Sieger" <powerhouse.linux@xxxxxxxxx>]

As a workaround, would PCI passthrough of a network adapter help?

On Monday, March 10, 2014 1:16 PM, Wei Liu <wei.liu2@xxxxxxxxxx> wrote:

Xen network driver is biased towards bulk transfer, so the usecase here
doens't quite fit.

On Tue, Mar 04, 2014 at 10:30:31PM +0200, NiX wrote:
> Hi. I've posted about this many months ago. Well I wanted to test the
> latest 4.3.x series XEN but unfortunately it seems the issue is still
> there.
>
> The issue is CPU usage when I zmap on a PV guest. I only used 10Mbs:
>
> zmap -p 80 -B 10M -o results.txt --interface=eth0
> --blacklist-file=/etc/zmap/blacklist_non_us.conf
>
> Mar 04 22:15:20.108 [INFO] zmap: started
> Mar 04 22:15:20.108 [INFO] zmap: zmap selected output module: csv
>  0:01 0%; send: 14914 14.8 Kp/s (14.3 Kp/s avg); recv: 375 374 p/s (359
> p/s avg); drops: 0 p/s (0 p/s avg); hits: 2.51%
>  0:02 0%; send: 29775 14.9 Kp/s (14.6 Kp/s avg); recv: 826 450 p/s (404
> p/s avg); drops: 0 p/s (0 p/s avg); hits: 2.77%
>  0:03 0%; send: 44664 14.9 Kp/s (14.7 Kp/s avg); recv: 1279 452 p/s (420
> p/s avg); drops: 0 p/s (0 p/s avg); hits: 2.86%
>  0:04 0%; send: 59563 14.9 Kp/s (14.7 Kp/s avg); recv: 1718 438 p/s (424
> p/s avg); drops: 0 p/s (0 p/s avg); hits: 2.88%
>
> So a 10Mbps & ~15k packets/second makes netback use 90% of CPU. Check
> attachment for more details. Dom0 is dual XEON X5450 and it's on 1Gbps
> network.
>
> I've verified zmap on bare metal and I can easily do 450k packets/second
> (300Mbps upload) and it uses about 150% CPU (1.5 cores) on Xeon CPU
> E3-1230 V2
>
> My problem is that I would like to virtualize few VM's that runs zmap on
> that Xeon CPU E3-1230 V2 machine but I am unable to do so because I can
> barely use 10mbps due to above issue.
>
> The kernel on Dom0 is:
> Linux nix 3.2.55-grsec #2 SMP Tue Mar 4 14:49:29 EET 2014 x86_64 GNU/Linux
>

Nonetheless I would suggest upgrading to 3.12+ Dom0 kernel as there's a
new model to handle guest TX which might help a little bit.

> PS. I've of course tweaked sysctl settings to support such gigantic
> packets/second so the kernel on bare metal is not the issue. I am not
> saying XEN is bad but CPU usage is way too high versus packets/second
>
> Has anyone tested similarly on VMware?


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


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

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