Mike Gerdts wrote: > On Thu, Apr 10, 2008 at 3:25 AM, Dan Price <[EMAIL PROTECTED]> wrote: >> # ptime pkg image-update (Null update) >> real 15.082 >> user 12.764 >> sys 0.229 >> >> Just over 6 minutes, or about 4.5MB/s = 36Mb throughput-- not too >> shabby. And clearly, there's room to squeeze that down further, >> although some of the optimizations will likely become more challenging. >> >> Anyway, this gives an approximate idea of how fast we might be able to >> provision enterprise systems in the data center, in the near future. >> Indeed, it could be even faster in a datacenter with gigabit links. > >>From looking at why it was taking so long to build a repository on my > Ultra 2, I could see that it was spending 90+% of its time in > compression routines (DTraceToolikt hotuser). When packages are > installed the files are decompressed at the client side, right? If > they are decompressed on the server side, the same issues will apply > for CMT repository servers. > > On a Niagara 2 box, created a /tmp/sbin.tar.gz of /usr/sbin, then used > "gzcat sbin.tar.gz> /dev/null" and found it could could decompress > this data at 5.9 MB/s. This would mean that on your 1990's LAN you are > pushing 76% of the data that one thread of a current system can > handle. > > Has any thought gone into making pkg multi-threaded? >
Yes. Right now, though, we're trying more to get the basics working well. Because Python's threading support is somewhat rudimentary compared to what we're used to, and because we want to explore alternative transports such as peer-to-peer, etc, I'm more inclined to split off the download and uncompress portions completely. When we get There's also the need for local caches of files so that repeated zone installations don't always download from the repository; this sort of caching is also very handy for doing installs over wireless, Comcast, and other networks of dubious reliability/integrity. Since we retrieve files by hash, and the hash will be easily locally verifiable against the signed manifests, our transports need not be all SSL to the mothership. - Bart -- Bart Smaalders Solaris Kernel Performance [EMAIL PROTECTED] http://blogs.sun.com/barts "You will contribute more with mercurial than with thunderbird." _______________________________________________ pkg-discuss mailing list [email protected] http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
