Open Stack

Hi, all

We have a benchmarking for openstack API, we found the performance is not
very well. For instance, with 3000 flavors in database, get all flavors API
TPS
is 4 under 10 concurrency. With 1000 images in glance, get all images API
TPS is
roughly 3. The hardware is Intel(R) Xeon(R) E5640 @ 2.67GHz, 48G memory.

As we know the most time-consuming in source code is ORM, so an obvious
 solution
 is do cache in nova-conductor, but nova-api does not get objects from
nova-conductor.

The easy way I can think of is inserting a cache system in API, maybe a
WSGI middleware
in api-paste.ini for better reusability. Then we can insert cache system in
any WSGI pipeline
bases application, like glance, neutron, etc. The cache system is policy
based. So different
API can have different expiry. For example, if flavors not changes often,
we can set expiry
of get flavors API to a bigger number. And instances changes often, the
expiry maybe 1
second. The backend of cache varies, memory based, file based, also a
dummy. The key
of cache is a mixture of URL, querystring, and headers, and only work for
"GET", "HEAD"
requests.

Make sense? Or another solution?
-- 
Best regards,
​TT
Gao
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140707/3ae1a100/attachment.html>