That helped -- using tempauth instead of keystone improved performance
significantly on the openStack cluster.
However performance is still slower compared to my Swift only setup where
commands are sent directly to the swift node instead of going thru the
controller node in the openstack cluster.
How does controller and swift node communications affect swift performance?
I've also noticed that all objects are stored on the 2nd swift node and not the
1st on the openstack cluster. I'm wondering if that could also be a factor in
slow performance.
Keystone:
TOTAL
Count: 50 Average requests per second: 9.2
min max avg std_dev 95%-ile
Worst latency TX ID
First-byte latency: 0.067 - 2.513 0.390 ( 0.604) 1.948 (all
obj sizes) txae75691d37d544b4ac0cfe3b8cba7f38
Last-byte latency: 0.067 - 3.337 0.430 ( 0.695) 1.997 (all
obj sizes) txdcedb82227654b338daa85751f6d1232
First-byte latency: 0.070 - 2.513 0.542 ( 0.749) 2.255 (
tiny objs) txae75691d37d544b4ac0cfe3b8cba7f38
Last-byte latency: 0.070 - 2.514 0.468 ( 0.659) 1.997 (
tiny objs) txae75691d37d544b4ac0cfe3b8cba7f38
First-byte latency: 0.067 - 1.884 0.251 ( 0.382) 0.695 (
small objs) tx2ceec827f3304530b01a0d5993eea2e8
Last-byte latency: 0.067 - 3.337 0.385 ( 0.732) 1.884 (
small objs) txdcedb82227654b338daa85751f6d1232
Tempauth:
Count: 50 Average requests per second: 65.7
min max avg std_dev 95%-ile
Worst latency TX ID
First-byte latency: 0.006 - 0.073 0.014 ( 0.015) 0.055 (all
obj sizes) tx69bf033a246645808b2c6a280e334f15
Last-byte latency: 0.006 - 0.248 0.047 ( 0.070) 0.198 (all
obj sizes) txb8cf5dc0ce264eb08a1e05edbbf5a40f
First-byte latency: 0.006 - 0.073 0.017 ( 0.020) 0.072 (
tiny objs) tx69bf033a246645808b2c6a280e334f15
Last-byte latency: 0.006 - 0.248 0.053 ( 0.072) 0.195 (
tiny objs) txb8cf5dc0ce264eb08a1e05edbbf5a40f
First-byte latency: 0.006 - 0.026 0.010 ( 0.005) 0.026 (
small objs) tx65d1fd4b6ae049bb902442ac4c28ffe9
Last-byte latency: 0.006 - 0.218 0.040 ( 0.066) 0.198 (
small objs) txbfd6ebc74ed04068affd17c123572a44
Swift Only:
TOTAL
Count: 50 Average requests per second: 397.0
min max avg std_dev 95%-ile
Worst latency TX ID
First-byte latency: 0.003 - 0.007 0.005 ( 0.001) 0.006 (all
obj sizes) None
Last-byte latency: 0.003 - 0.046 0.008 ( 0.009) 0.029 (all
obj sizes) None
First-byte latency: 0.003 - 0.007 0.005 ( 0.001) 0.007 (
tiny objs) None
Last-byte latency: 0.003 - 0.046 0.008 ( 0.010) 0.027 (
tiny objs) None
First-byte latency: 0.004 - 0.006 0.005 ( 0.001) 0.006 (
small objs) None
Last-byte latency: 0.004 - 0.043 0.008 ( 0.009) 0.029 (
small objs) None
From: Chmouel Boudjnah [mailto:[email protected]]
Sent: Tuesday, July 09, 2013 8:22 AM
To: OpenStack Development Mailing List
Subject: Re: [openstack-dev] Swift debugging / performance - large latencies
seen.
On Tue, Jul 9, 2013 at 2:40 PM, Snider, Tim
<[email protected]<mailto:[email protected]>> wrote:
I have 2 openstack clusters running the Folsom release with multiple Swift
nodes. I also have a small setup that is running only Swift with a single node.
I'm noticing very large Swift I/O latencies (seconds long) on the openstack
clusters - ssbench output snippet is below. Performance is approximately
identical on the openstack clusters. The Swift only cluster performs much
better.
Keystone performance can be pretty awful unless you are using something else
than the default WSGI container configuration (single process eventlet I
think). I would suggest you try to run it under apache with multiple process.
See the dicussion at last summit about Keystone performance here :
https://etherpad.openstack.org/havana-keystone-performance
Chmouel.
_______________________________________________
OpenStack-dev mailing list
[email protected]
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
