I saw (reading my referrer logs :-) people were interested in distcc with Wine, so I thought I would add it to distcc's performance benchmark.
Across three machines I saw a 2.58x speedup -- details are attached. The machines were all 1GB, 1700MHz Pentium IV with a 100Mbps switched connection. Building on just one took 22m57s, and across all three took 8m53s. There is probably still some scheduling inefficiency here, which I hope to reduce over the next few months.
I'm not sure what was happening with the person who saw little change. Perhaps they had distcc misconfigured, or perhaps they hit a bug. Either way I would be interested in hearing the details ($DISTCC_HOSTS, etc) and seeing if I can help fix it.
I hope you find ccache and distcc useful. Wine is a cool project.
Please cc me on replies.
Martin Pool wrote:
I saw (reading my referrer logs :-) people were interested in distcc with Wine, so I thought I would add it to distcc's performance benchmark.
Across three machines I saw a 2.58x speedup -- details are attached. The machines were all 1GB, 1700MHz Pentium IV with a 100Mbps switched connection. Building on just one took 22m57s, and across all three took 8m53s. There is probably still some scheduling inefficiency here, which I hope to reduce over the next few months.
Cool!
I'm not sure what was happening with the person who saw little change. Perhaps they had distcc misconfigured, or perhaps they hit a bug. Either way I would be interested in hearing the details ($DISTCC_HOSTS, etc) and seeing if I can help fix it.
I think the issue was that I did not have three identical machines; the speedup from distributing to three hosts was minimal compared to just building on the fast machine.
I hope you find ccache and distcc useful. Wine is a cool project.
Thanks for working on distcc et al! BTW, is gecc ready for prime time yet? Presumably it would be a win over either ccache or distcc? - Dan
On 8 Jan 2003, Dan Kegel dank@kegel.com wrote:
I think the issue was that I did not have three identical machines; the speedup from distributing to three hosts was minimal compared to just building on the fast machine.
I think the scheduler doesn't handle the case of heterogenous machines as well as it might. I'm going to try to fix that.
You should make sure that the fastest machines are listed first in DISTCC_HOSTS.
Thanks for working on distcc et al! BTW, is gecc ready for prime time yet?
My impression is that it is still experimental, but I don't work on it so I can't really say.
Presumably it would be a win over either ccache or distcc?
Integrated distributed caching and compilation is probably best in the long term. My approach was to get "small sharp" tools really solid first before tackling integration -- correctness is more important than speed. I've discovered a number of interesting gcc "behaviours" along the way and they might have been harder to isolate with a more complex tool.
Martin Pool wrote:
On 8 Jan 2003, Dan Kegel dank@kegel.com wrote:
I think the issue was that I did not have three identical machines; the speedup from distributing to three hosts was minimal compared to just building on the fast machine.
I think the scheduler doesn't handle the case of heterogenous machines as well as it might. I'm going to try to fix that.
You should make sure that the fastest machines are listed first in DISTCC_HOSTS.
OK, so if I'm running the build on the fast host, I should include 'localhost' first in DISTCC_HOSTS.
But what about SMP machines? If the fast machine is a dual processor box, and the others are all single-processor, should I put localhost twice in DISTCC_HOSTS?
That's probably part of the reason I got such a poor speedup -- I had no idea how to configure distcc well in the first place, let alone configuring it to deal with SMP systems. - Dan
-
Dan Kegel -
Martin Pool