[openstack-dev] [Fuel] 10Gbe performance issue.
piotr.skamruk at intel.com
Wed Jan 21 10:53:42 UTC 2015
On Tue, 2015-01-20 at 17:41 +0100, Tomasz Napierala wrote:
> How this was measured? VM to VM? Compute to compute?
iperf between compute/ceph-compute/ceph nodes.
> In any case, what is your deployment configuration, especially VLAND or GRE, networking gear, etc.
We have almost default setup from clean fuel 6.0, vlan based.
We are doing iperf based on 82599ES 10-Gigabit SFP+ interfaces (used in
our setup only for storage network) connected by fibre through Arista
Results on centos 6.5 deployed from fuel:
0.0-10.0 sec 3.09 GBytes 2.65 Gbits/sec
Results on centos 6.5 from official site:
0.0-10.0 sec 10.9 GBytes 9.38 Gbits/sec
In both cases tests were done with default tcp window size - 19.3KByte.
In both cases default mtu was with value 1500.
Commands used in test (after adding iperf tool from rpm):
on one node:
iperf -s -p 8777
on other node:
iperf -c ip_address_of_server_in_storage_network -p 8777 -t 10
(port 8777 is one of passed as ACCEPT in iptables setup from fuel)
Differences are in kernel (from centos or from mirantis), values
in /etc/sysctl.conf (plain or from mirantis) and probably in cgroups
Probably in ~30 minutes we also will have results on plain centos with
mirantis kernel, and on fuel deployed centos with plain centos kernel
(2.6.32 in both cases, but with different patchset subnumber).
More information about the OpenStack-dev