LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: [lvs-users] Fwd: Re: LVS throughput limit (450k pps)?

To: Patrick Schreurs <pats@xxxxxxxxx>
Subject: Re: [lvs-users] Fwd: Re: LVS throughput limit (450k pps)?
Cc: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
From: Alexandre Cassen <acassen@xxxxxxxxxx>
Date: Tue, 15 Jan 2008 11:46:13 +0100
Hi Patrick,

> > Handling loadbalancing in softirq corner is, IMHO, not the best place
> > for performance. Softirq handling design become a performance bottlneck
> > catalysed by the democratization of SMP (multi-core) box. There is
> > alternative design but need to find time to work fulltime on this (I
> > have this on my TODO).
> 
> The speed of a single core is just too slow. Balancing the interrupts
> over multiple CPU cores gives a performance penalty.
> 
> What about MSI-X with multiqueue (like intel chipset 82575). Will that
> help. I haven't found a nic with that chipset yet, but that's a different
> story.

multi-rx/tx queue is great if supported by NIC driver you are using. for
example the ixgb driver is supporting multiqueue handling. But with a
simple rework of the softirq design you can have a huge performance
gain. I mean, skb are coming from netdev and directly queued into
multi-lvs-queue which job is to drive loadbalancing decision + skb
gardening, this queue is fetched by dedicated kthread, so your netdev
only fill skb into the queue while kthread unqueue/handle. With this
simple code extension you will have a nice gain. IMHO, multiqueue is
great design and is legitimate by NIC bandwidth, I mean, for 10GE NIC.

Regards,
Alexandre



<Prev in Thread] Current Thread [Next in Thread>
  • Re: [lvs-users] Fwd: Re: LVS throughput limit (450k pps)?, Alexandre Cassen <=