Slightly OT, but do you see TCP Fast Open as something that will help reduce
your pps rate? In the past, you mentioned email being a big part of your load,
so I am guessing that your connections are longer lived and that it will not
be as much of a factor as some web sites that have many short lived connections.
I doubt you have tried this (I know Dell sells AMD boxes but I am not sure if
any Dell customers buy them) but have you ever seen the performance
differences for an AMD based box in your high load LVS director work load vs
the Intels as far as softirq, etc?
Thanks for sharing.
On 3/7/13 11:17 AM, lvs@xxxxxxxxxx wrote:
> I am doing 500k packets per second in each direction which is in the same
> kind of order you are looking for.
> I am running this through a Dell r710 with two Intel X520-DA2 NICs, giving
> me four 10Gb ports, 2 external and 2 internal. Each pair of ports is
> configured in an active-passive bond. I am passing the return traffic
> through the load balancer and doing firewalling as well as load balancing.
> The r710 has the top CPUs Intel offered at the time (3.3GHz? 6 cores + HT
> per socket for 24 core total). I am seeing the businest core at 50% with
> this workload so there is more capacity there.
> I have some r720s with the top 8 core CPUs (2.9GHz?) which are looking
> much better for the job, but I haven't switched the main load in any DC to
> an r720 yet so this is based on how they handle 100-200kpps.
> In the past I split the firewalling to a box in front of a pure load
> balancer (without iptables modules loaded) and I saw the load split evenly
> across the two servers, so if you don't need firewalling it should go even
> faster. With firewalling I find CPU load drops significanly with IPVS-NFCT
> turned on.
> On Thu, 7 Mar 2013, Rinaldo Digiorgio wrote:
>> I am new to this alias. I have been looking at various load balancer
>> configurations, Zeus, HA Proxy, Balance/NG and AWS to determine if there is
>> some solution that gives me 400K small requests per second say a get of an
>> index page with the possibility of routing the response directly to the
>> user. Has anyone come close with some of the faster Linux boxes with 10GE
>> cards? I can get 100K requests with HA proxy and was unable to get
>> Balance/NG configured, it has a very odd configuration file where they
>> reverse the meanings of words and expect users to be able to deal with it.
>> I guess I am just asking for any numbers that anyone can share on what rates
>> they have achieved.
>> Thanks for the time.
>> Please read the documentation before posting - it's available at:
>> 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
> Please read the documentation before posting - it's available at:
> 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
Please read the documentation before posting - it's available at:
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