An alternative is the tree-based data structure which has its own
advantages and disadvantages. The current structure is very good, as
there is no overhead in updating the ruleset and leverages the current
limitations on wildcard bitpattern. However, its performance depends on
how the rules have been defined. If they are scattered in different
wildcard bitpattern, the performance is low.
In order to reduce the CPU usage, reducing the number of rules is not
effective comparing to minimizing the number of unique wildcard bits.
For example reducing the bitpatterns from 900 to 400 or the number of
rules from 240k to 1.2k for 600 bitpatterns both reduce the CPU usage
from 100% to 50% with the same number of new flows.
On 4/16/2012 8:14 PM, Ben Pfaff wrote:
On Mon, Apr 16, 2012 at 04:23:50PM -0700, Masoud Moshref Javadi wrote:
The performance of OVS new flow setup (at least 1.4) mostly depends
on the number of unique wildcard patterns in the rules it saves
because the classifier goes linearly through the hashsets to find a
match. For example it can support about 10k rules in 900 unique
wildcard patterns with 100% of a CPU core allocation if we have 1500
new flows per second.
What alternative classifier structure would you suggest instead?
--
Masoud Moshref Javadi
Computer Engineering PhD Student
Ming Hsieh Department of Electrical Engineering
University of Southern California
_______________________________________________
openflow-discuss mailing list
openflow-discuss@lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
