Re: [RFC PATCHv2 04/10] gitweb: Use Cache::Cache compatibile (get, set) output caching

On Wed, 10 Feb 2010, Petr Baudis wrote:

On Wed, Feb 10, 2010 at 02:12:24AM +0100, Jakub Narebski wrote:

quoted

On Tue, 9 Feb 2010 at 11:30 +0100, Jakub Narebski wrote:

quoted

The cache_fetch subroutine captures output (from STDOUT only, as
STDERR is usually logged) using either ->push_layer()/->pop_layer()
from PerlIO::Util submodule (if it is available), or by setting and
restoring *STDOUT.  Note that only the former could be tested reliably
to be reliable in t9503 test!

Scratch that, I have just checked that (at least for Apache + mod_cgi,
but I don't think that it matters) the latter solution, with setting
and restoring *STDOUT doesn't work: I would get data in cache (so it
can be restored later), but instead of output I would get Internal Server
Error ("The server encountered an internal error or misconfiguration and
was unable to complete your request.") without even a hint what the
problem was.  Sprinkling "die ...: $!" didn't help to catch this error:
I suspect that the problem is with capturing.

So we either would have to live with non-core PerlIO::Util or (pure Perl)
Capture::Tiny, or do the 'print -> print $out' patch...

All the magic methods seem to be troublesome, but in that case I'd
really prefer a level of indirection instead of filehandle - as is,
'print (...) -> output (...)' ins. of 'print (...) -> print $out (...)'
(or whatever). That should be really flexible and completely
futureproof, and I don't think the level of indirection would incur any
measurable overhead, would it?

First, it is not only 'print (...) -> print $out (...)'; you need to
do all those:

   print  <sth>            ->  print  $out <sth>
   printf <sth>            ->  printf $out <sth>
   binmode STDOUT, <mode>  ->  binmode $out, <mode>

Second, using "tie" on filehandle (on *STDOUT) can be used also for 
just capturing output, not only for "tee"-ing; what's more to print
while capturing one has to do extra work.  It is quite similar to
replacing 'print (...)' with 'output (...)' etc., but using
tie/untie doesn't require large patch to gitweb.

Third, as you can see below tie-ing is about 1% slower than using
'output (...)', which in turn is less than 10% slower than explicit
filehandle solution i.e. 'print $out (...)'... and is almost twice
slower than solution using PerlIO::Util

Benchmark: timing 50000 iterations of output, perlio, print \$out, tie *STDOUT...
     output: 1.81462 wallclock secs ( 1.77 usr +  0.00 sys =  1.77 CPU) @ 28248.59/s (n=50000)
     perlio: 1.05585 wallclock secs ( 1.03 usr +  0.00 sys =  1.03 CPU) @ 48543.69/s (n=50000)
print \$out: 1.70027 wallclock secs ( 1.66 usr +  0.00 sys =  1.66 CPU) @ 30120.48/s (n=50000)
tie *STDOUT: 1.82248 wallclock secs ( 1.79 usr +  0.00 sys =  1.79 CPU) @ 27932.96/s (n=50000)
               Rate tie *STDOUT      output print \$out      perlio
tie *STDOUT 27933/s          --         -1%         -7%        -42%
output      28249/s          1%          --         -6%        -42%
print \$out 30120/s          8%          7%          --        -38%
perlio      48544/s         74%         72%         61%          --

Benchmark: running output, perlio, print \$out, tie *STDOUT for at least 10 CPU seconds...
     output: 10.7199 wallclock secs (10.53 usr +  0.00 sys = 10.53 CPU) @ 28029.63/s (n=295152)
     perlio: 11.2884 wallclock secs (10.46 usr +  0.00 sys = 10.46 CPU) @ 49967.11/s (n=522656)
print \$out: 10.5978 wallclock secs (10.43 usr +  0.00 sys = 10.43 CPU) @ 30318.79/s (n=316225)
tie *STDOUT: 11.3525 wallclock secs (10.68 usr +  0.00 sys = 10.68 CPU) @ 27635.96/s (n=295152)
               Rate tie *STDOUT      output print \$out      perlio
tie *STDOUT 27636/s          --         -1%         -9%        -45%
output      28030/s          1%          --         -8%        -44%
print \$out 30319/s         10%          8%          --        -39%
perlio      49967/s         81%         78%         65%          --
need

Attached there is script that was used to produce those results.
-- 
Jakub Narebski
Poland