Testing box available.
Kris Kennaway
kris at FreeBSD.org
Thu Feb 14 10:36:36 UTC 2008
Julian Elischer wrote:
> my memory is that ohk changed make to have a fifo with N (as in -j N)
> tokens in it and all child makes inherrit this fifo
> (or get it's name from an environment variable or something)
> and can only spawn extra makes if they can get a token.
> When they have finished their work they pu the token back into the
> fifo..
>
> My memories of this may be somewhat inaccurate however.
I think this is correct but it only helps to some extent. Large parts
of the buildworld have no or low parallelism still (e.g. the whole 'make
depend' phase).
'make buildworld' is not a good choice for testing parallelism. Running
a "useful workload" is more interesting :) One of three things will happen:
1) It may scale well to 16 CPUs.
2) The scheduler may limit performance if your workload depends heavily
on CPU locality. This is the case with the Intel 16-core system we have
been using. The scheduler needs to gain knowledge of the CPU core
topology in order to avoid making bad decisions about migrating
processes between CPU cores that are far apart. Jeff is working on this.
3) Your workload may not scale because of a limiting factor in userland
or in the kernel. If this is a limitation in FreeBSD and not the
application then it is interesting.
Actually it would be interesting to repeat some of our "standard tests"
on the barcelona system to see how it compares to the intel one.
Kris
More information about the freebsd-current
mailing list