[3.10.22+][3.12][performance regression] - huge softirq increase on netfilter

[3.10.22+][3.12][performance regression] - huge softirq increase on netfilter

[Nikolay Ulyanitsky]

Hello


3.10.22 has new 2 patches that increase softirq from 0.0% to 15.0% in
comparison with 3.10.21 on my software routers:

78ab8cb      netfilter: nf_conntrack: use RCU safe kfree for conntrack
extensions
873be93      ipc/sem.c: synchronize semop and semctl with IPC_RMID


Screenshots:
Server1 (ipt_NETFLOW):     http://s12.postimg.org/5eprc03tp/vega.png
Server2 (NAT):    http://s27.postimg.org/swj3he8dv/mimosa.png


Server1:
3.10.21: softirq = 0%
3.10.22+, 3.12.4: softirq = 15%
3.10.22+, 3.12.4 WITHOUT 78ab8cb, WITH 873be93: softirq = 1%
3.10.22+, 3.12.4 WITHOUT 78ab8cb, WITHOUT 873be93: softirq = 0%


Hardware:
2 x Xeon 3.2GHz, Intel SE7520, 4 x Intel 82576 Gigabit Network

CentOS 6.5 x86_64


Kernel configs:
3.12.4: http://pastebin.com/ugTWFeJg
3.10.22: http://pastebin.com/N8z4UAV5


Please fix/revert these patches.


Thanks


–

With best regards,

Nikolay Ulyanitsky

Re: [3.10.22+][3.12][performance regression] - huge softirq increase on netfilter

[Paul E. McKenney]

On Wed, Dec 11, 2013 at 02:01:29PM +0200, Nikolay Ulyanitsky wrote:
> Hello
> 
> 
> 3.10.22 has new 2 patches that increase softirq from 0.0% to 15.0% in
> comparison with 3.10.21 on my software routers:
> 
> 78ab8cb      netfilter: nf_conntrack: use RCU safe kfree for conntrack
> extensions

This one fixes a use-after-free bug, so reverting it might be unwise.
However, it looks like it should only be affecting networking connection
tracking.  Is that an important part of your workload?

Also, are you checking overall CPU consumption?  One effect of this
patch is to shift overhead from process level to softirq, so please check
process-level CPU consumption to make sure you have a fair comparison.

> 873be93      ipc/sem.c: synchronize semop and semctl with IPC_RMID

This one also fixes some bugs.  But I don't immediately see why it
should increase softirq CPU consumption.

						Thanx, Paul

> Screenshots:
> Server1 (ipt_NETFLOW):     http://s12.postimg.org/5eprc03tp/vega.png
> Server2 (NAT):    http://s27.postimg.org/swj3he8dv/mimosa.png
> 
> 
> Server1:
> 3.10.21: softirq = 0%
> 3.10.22+, 3.12.4: softirq = 15%
> 3.10.22+, 3.12.4 WITHOUT 78ab8cb, WITH 873be93: softirq = 1%
> 3.10.22+, 3.12.4 WITHOUT 78ab8cb, WITHOUT 873be93: softirq = 0%
> 
> 
> Hardware:
> 2 x Xeon 3.2GHz, Intel SE7520, 4 x Intel 82576 Gigabit Network
> 
> CentOS 6.5 x86_64
> 
> 
> Kernel configs:
> 3.12.4: http://pastebin.com/ugTWFeJg
> 3.10.22: http://pastebin.com/N8z4UAV5
> 
> 
> Please fix/revert these patches.
> 
> 
> Thanks
> 
> 
> –
> 
> With best regards,
> 
> Nikolay Ulyanitsky
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>