nick at ccl4.org
Tue Feb 25 16:45:43 GMT 2014
On Tue, Feb 25, 2014 at 05:37:26PM +0100, Abigail wrote:
> On Tue, Feb 25, 2014 at 03:59:21PM +0000, Nicholas Clark wrote:
> > IIRC booking.com thought that they got roughly a 30% CPU drop in going from
> > vendor supplied 5.8.5 to vanilla 5.14.mumble built from source.
> > They never did blog this though :-(
> Yeah, but from what I heard, most of the gain was in the custom compilation
> parameters. I wasn't involved in the process though, so I may be mistaken.
A gain is still a gain. Anyone else can similarly recompile with more
suitable options (or a newer compiler, or ...)
eg if you don't need threads, then you don't need to pay the penalty of the
multiple interpreter indirection ithreads adds. And I think pretty much
every OS-supplied perl is paying this, so building your own is an instant
free win there.
IIRC v5.14 had a runloop optimisation which reduces typical CPU usage by
maybe 5%, so there's some real gain, even with everything else held constant.
(And v5.10.1 and later have O(1) ISA lookup rather than O(n), which I believe
was a big speedup for PPI)
Vendor supplied perl is going to be slower than one you compile yourself.
Or at least, it won't be faster.
> We're moving to 5.18.something very soon -- perhaps someone will blog
> about that.
Would be cool if they did.
More information about the london.pm