On Thu, May 12, 2011 at 03:40, Mark Phippard <markp...@gmail.com> wrote: > I ran the benchmark tests again using latest trunk for server and > client. The numbers are interesting. See them here so you can see > table formatting: > > https://ctf.open.collab.net/sf/wiki/do/viewPage/projects.csvn/wiki/HTTPv2 > > Using Neon, performance is improved across the board. In some cases, > substantially. The number of HTTP requests were cut from 11,723 down > to 5,415! > > The story using Serf is not as good. There are a few places where it > is fastest, namely merge. But there are other cases where it is > dramatically slower. The number of HTTP requests with Serf is 80,990. > Looking at the numbers, it seems like Serf is slowest when it comes > to the areas where it issues all those GET requests, such as checkout > and update. On other operations it is more inline with Neon. Maybe > there is more that could be done to speed up those areas or maybe you > need to get the server under high load for the benefits of the Serf > approach to manifest? FWIW, I monitored the CPU usage of the server > during the run. The utilization was low throughout the runs but > definitely higher when Serf was used. That said, I could also see it > using both CPU cores when Serf was used, but not with Neon. > > When you are in Berlin next week I hope you can revisit the discussion > about whether Serf should be our default. Right now, unless there are > more improvements, I am going to advise that we patch our CollabNet > binaries so that Neon is the default. We will of course still include > Serf so that it can be configured but I do not think it makes sense > for it to be the default if it is going to be considerably slower at > most operations. > We (VisualSVN) are considering doing the same in our packages.
Another option is to implement non-skelta mode in ra_serf, but Greg has objections for this way. I hope we can discuss this in Berlin next week. -- Ivan Zhakov
