Hi All, We have been working on MB2 performance improvements, based on Disrupter and then batching Cassandra calls and removing some non-critical cassandra calls (e.g. message counters).
Following are some initial results 1) With one MB Node, With cached senders, I got about 13K TPS in new perf machines publishing and about 2500 TPS at the receiving side. 2) With one MB Node, With new senders per message (this make sure message is written to the disk), about 2500 TPS at publisher and 2500 TPS at receiver side. (Earlier we got about 1500 TPS before changes ). In the released version, if we send about a Million message very fast .. memory get accumulated and system goes OOM. With disrupter changes this limit is gone. We tested with 3M messages back to back without any trouble. Receiving side bottleneck is caused by Ack sent per message. AQMP support batched Acks, but need to work on that. We have not tested multiple broker scenarios, but the bottleneck is not Cassandra any more. So we think it should scale well. However, these changes need to be carefully reviewed to make sure it does not break anything major. If all is well, we can do a point release with performance improvements. --Srinath -- ============================ Srinath Perera, Ph.D. Director, Research, WSO2 Inc. Visiting Faculty, University of Moratuwa Member, Apache Software Foundation Research Scientist, Lanka Software Foundation Blog: http://srinathsview.blogspot.com/ Photos: http://www.flickr.com/photos/hemapani/ Phone: 0772360902
_______________________________________________ Architecture mailing list [email protected] https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
