Good aproach. It's like a stack (LIFO) of sched_connections BUT I'd prefer a linked list, because it's simpler. You could use just a "free list" and not two arrays (stack and queue). When a connection is closed his sched_connection is returned to the "free list" (head).
Another issue in mk_scheduler is mk_sched_get_connection(). This function is called from mk_conn_write() and mk_sched_remove_client(). The mk_sched_get_connection()'s complexity is O(work_capacity), is used two times at least in connection life when it could be avoid completely. epoll_wait returns a event array and Monkey uses the socket fd as epoll_event data. That's wrong decision!, epoll_event data should be the sched_connection and NOT the socket fd. It's possible to improve it, but need hard work. El 26-03-2012, a las 1:22, Eduardo Silva escribió: > Hi, > > thanks for the patch. Looking with valgrind seems to be optimized a > little bit, screenshot here: > > http://edsiper.linuxchile.cl/sched_optimization_001.png > > without optimization mk_sched_register() takes 0.40 for 5000 calls, > the same situation but for an optimized code takes 0.36. Its an > improvement. > > Dave, Zeus and Max, what do you think about the patch ? > > cheers, > > > On Sun, Mar 25, 2012 at 9:43 PM, Mahesh Gondi <[email protected]> wrote: >> Hi all, >> >> I made some changes to mk_scheduler.c. First I will explain in brief what I >> did before the results. >> >> In mk_scheduler.c , the mk_sched_register_client serves the purpose of >> adding new client requests to the worker thread queue(everything discussed >> here happens in the thread context). Adding was done by iterating over the >> queue to looking for an available spot to be inserted. When the load on >> server is at near max, then this insertion cost rises to O(work_capacity). >> >> Instead I maintained free spots on the queue(list of client requests >> received), in a simple array of size (work_capacity+1) with each element >> pointing to an index in queue(first element kept a count of number of free >> spots available). Array(arr) contains free spots as pointed by the index >> values stored at the position from 1 to arr[0]. Insertion now only takes a >> constant time. Hence this has contributed in running monkey a bit cheaper. >> Similar modifications are in progress, should help monkey run more and more >> faster . :) >> >> Below are the results >> >> Output I got for running with "siege -c 300 -t 30S 127.0.01:2001", >> >> //WITH CONSTANT TIME INSERTION >> Transactions: 18051 hits >> Availability: 100.00 % >> Elapsed time: 29.96 secs >> Data transferred: 23.48 MB >> Response time: 0.00 secs >> Transaction rate: 602.50 trans/sec >> Throughput: 0.78 MB/sec >> Concurrency: 2.30 >> Successful transactions: 18051 >> Failed transactions: 0 >> Longest transaction: 0.23 >> Shortest transaction: 0.00 >> >> ============================================ >> >> //EARLIER >> Transactions: 17711 hits >> Availability: 100.00 % >> Elapsed time: 30.01 secs >> Data transferred: 23.04 MB >> Response time: 0.00 secs >> Transaction rate: 590.17 trans/sec >> Throughput: 0.77 MB/sec >> Concurrency: 1.18 >> Successful transactions: 17711 >> Failed transactions: 0 >> Longest transaction: 0.17 >> Shortest transaction: 0.00 >> >> i had taken output for each case just after a fresh restart. Reason for only >> ~600 trans/sec is that it was run ec2 t1.small instance. >> >> Thanks & Regards, >> mahesh gondi >> >> _______________________________________________ >> Monkey mailing list >> [email protected] >> http://lists.monkey-project.com/listinfo/monkey >> > > > > -- > Eduardo Silva > http://edsiper.linuxchile.cl > http://www.monkey-project.com > _______________________________________________ > Monkey mailing list > [email protected] > http://lists.monkey-project.com/listinfo/monkey _______________________________________________ Monkey mailing list [email protected] http://lists.monkey-project.com/listinfo/monkey
