[pve-devel] [PATCH] add hugepages option v7

[Alexandre DERUMIER]

>>Thats impressive, but maybe a very special case? 64Bytes frame size is very
>>small, so
>>it does not really matter in real live?

for a virtual router, yes !  (think about a synflood attack for example)

Also, for latency, applications like redis really need it


>>ovs+dpdk is also slower than linux on other frame sizes... 

Where are you seeing that ?

AFAIK, ovs (without dpdk) and linux (bridge or routing) are both limited to around 1,5-2 mpps for 1core (any framesize).

Linux kernel devs are still working to improve this, but it's far from the speed of dpdk.
http://people.netfilter.org/hawk/presentations/NetDev1.1_2016/net_performance_BoF.pdf






----- Mail original -----
De: "dietmar" <dietmar at proxmox.com>
À: "aderumier" <aderumier at odiso.com>
Cc: "pve-devel" <pve-devel at pve.proxmox.com>
Envoyé: Lundi 20 Juin 2016 07:08:49
Objet: Re: [pve-devel] [PATCH] add hugepages option v7

> My DBA engineer have seen around 20% improvement on sqlserver workload vs 2MB 
> hugepages, 
> but I don't have a reproductible benchmark. (It's was stats on real workload 
> on a 1week stats summary) 
> 
> 
> Note that hugepage is needed for dpdk, 

sure, i know... 

> and if you want some dpdk benchmark results: 
> 
> https://download.01.org/packet-processing/ONPS1.5/Intel_ONP_Server_Release_1.5_Performance_Test_Report_Rev1.2.pdf 
> 
> 1cpu core 
> --------- 
> ovs : framesize:64Bytes result: 1,103,295 pps 
> ovs+dpdk : framesize:64Bytes result: 16,766,108 pps 
> 
> so, x16 

Thats impressive, but maybe a very special case? 64Bytes frame size is very 
small, so 
it does not really matter in real live? 

ovs+dpdk is also slower than linux on other frame sizes...