On 16 Jun 2010, at 12:49 PM, Joe Orton wrote:
The core output filter should coalesce "small" writes already and if
it's not doing that effeciently that's a bug. Are the "tiny" files
you
were seeing getting sendfile()d out bigger than AP_MIN_SENDFILE_BYTES?
That sounds like a case for bumping up AP_MIN_SENDFILE_BYTES.
Many of the packets consisted of a CRLF followed by an indentation
tab, and were only a few bytes long. The core output filter is too
late in the chain to make any difference, as the response had already
been turned by the chunking filter into hundreds of little chunks,
consisting of a heap bucket (chunk header), followed by a file bucket
(the original content), followed by an immortal bucket (trailing
CRLF), etc.
The fix was to simply buffer up the fragmented output directly after
the fragments where created, and let the rest of the filter stack step
in and work as efficiently as possible on as few buckets as possible.
Regards,
Graham
--
