LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: [lvs-users] RX Packet drops on high traffic LVS in DR setup

To: "LinuxVirtualServer.org users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] RX Packet drops on high traffic LVS in DR setup
From: Sashi Kant <sashi.kant@xxxxxxxxxxxxx>
Date: Fri, 24 Oct 2008 16:48:42 -0700
I am trying several different approaches to this problem.

1. Ethernet driver update - upgraded to kernel ethchnhalf with
bnx2-firmware , situation improves , however still dropping packets
during traffic spikes.

2. Broadcom cards sit on a PCI-X bridge off PCI Express bus, We tried
other add-on broadcom card and it did not help as these cards were on
same PCI-X bus.

3. Added PCI Express card and things looking better, very few drops when
traffic spikes suddenly.

4. Removed bonding, Cisco Ether Channel configuration and use PCIe
gigabit card from intel (conf 3). We are currently testing this config
and there are no packet drops at 50K pps.

I will update you when we find our optimal configuration.

IMHO Bus, NIC are huge players in this mix.

Thank you all for your kind insights.

-  

Thanks
-Sashi


On Fri, 2008-10-24 at 08:01 +0100, Graeme Fowler wrote:
> On Fri, 2008-10-24 at 04:50 +0000, MontyRee wrote:
> > I would like to know the maximal performance(PPS) of LVS too.
> > and you mean that 40kpps is the limit of the LVS?
> 
> No.
> 
> > anyone who tested maximal PPS performance?
> > any recommendation to improve the maximal PPS?
> 
> Better, bigger, faster hardware.
> 
> Off the top of my head:
> 
> Multiple CPUs
> Multiple NICs on separate PCI buses
> The fastest PCI bus and front-side bus you can get
> 
> Maybe, just maybe, AMD boards might be able to handle more interrupts -
> since every packet generates an interrupt [0] - because of their
> differing hardware architecture to Intel. That's a suggestion and not a
> scientifically proven statement!
> 
> [0] This may not be true on some hardware and some operating systems,
> since the packets can be aggregated if they're small and several sent up
> the bus to interrupt the CPU at the same time.
> 
> The problem is that there's almost no benchmarks available, because
> every single system in different.
> 
> Graeme
> 
> 
> 
> _______________________________________________
> LinuxVirtualServer.org mailing list - lvs-users@xxxxxxxxxxxxxxxxxxxxxx
> Send requests to lvs-users-request@xxxxxxxxxxxxxxxxxxxxxx
> or go to http://lists.graemef.net/mailman/listinfo/lvs-users



<Prev in Thread] Current Thread [Next in Thread>