> -----Original Message----- > From: n...@orbyte.nwl.cc [mailto:n...@orbyte.nwl.cc] On Behalf Of Phil > Sutter > Sent: Saturday, January 6, 2018 1:25 AM > To: Chris Mi <chr...@mellanox.com> > Cc: netdev@vger.kernel.org; gerlitz...@gmail.com; > step...@networkplumber.org; dsah...@gmail.com; > marcelo.leit...@gmail.com > Subject: Re: [patch iproute2 v6 0/3] tc: Add -bs option to batch mode > > Hi Chris, > > On Thu, Jan 04, 2018 at 04:34:51PM +0900, Chris Mi wrote: > > Currently in tc batch mode, only one command is read from the batch > > file and sent to kernel to process. With this patchset, we can > > accumulate several commands before sending to kernel. The batch size > > is specified using option -bs or -batchsize. > > > > To accumulate the commands in tc, client should allocate an array of > > struct iovec. If batchsize is bigger than 1, only after the client has > > accumulated enough commands, can the client call rtnl_talk_msg to send > > the message that includes the iov array. One exception is that there > > is no more command in the batch file. > > > > But please note that kernel still processes the requests one by one. > > To process the requests in parallel in kernel is another effort. > > The time we're saving in this patchset is the user mode and kernel > > mode context switch. So this patchset works on top of the current kernel. > > > > Using the following script in kernel, we can generate 1,000,000 rules. > > tools/testing/selftests/tc-testing/tdc_batch.py > > > > Without this patchset, 'tc -b $file' exection time is: > > > > real 0m15.555s > > user 0m7.211s > > sys 0m8.284s > > > > With this patchset, 'tc -b $file -bs 10' exection time is: > > > > real 0m13.043s > > user 0m6.479s > > sys 0m6.504s > > > > The insertion rate is improved more than 10%. > > Did you measure the effect of increasing batch sizes? Yes. Even if we enlarge the batch size bigger than 10, there is no big improvement. I think that's because current kernel doesn't process the requests in parallel. If kernel processes the requests in parallel, I believe specifying a bigger batch size will get a better result. > > I wonder whether specifying the batch size is necessary at all. Couldn't batch > mode just collect messages until either EOF or an incompatible command is > encountered which then triggers a commit to kernel? This might simplify > code quite a bit. That's a good suggestion.
-Chris > > Cheers, Phil
