[cvsnt] Re: wnt_fclose and FlushFileBuffers

[John Goehringer]

It's a snow day today - about 10 inches or so far so I'm testing
over a 376k DSL instead of the 100Mbit switched network.

The test consists of about 530 files, mostly Java source and xml, 3Mb
total size.

With Build 69 on the either the client or server, a cvs -z9 checkout takes
about 75 seconds. The same with Build 68 on both the client and server
takes 13 seconds.

Here are the individual results. The 1.3Ghz PIII is the production server,
so I did not test it with Build 69. Everything is running W2K, Pro on the
clients, Server on the servers.

Client 800Mhz PIII 68 Server 1.3Ghz PIII 68

real    0m12.651s

Client 800Mhz PIII 69 Server 1.3Ghz PIII 68

real    1m16.362s

Client 800Mhz PIII 69 Server 400Mhz PII 69

real    1m14.423s

Client 800Mhz PIII 69 Server 400Mhz PII 68

real    1m16.262s

Client 800Mhz PIII 68 Server 400Mhz PII 68

real    0m12.237s


"Tony Hoyle" <tmh at nodomain.org> wrote in message
news:b1umuh$hm9$1 at sisko.nodomain.org...
> John Goehringer wrote:
>
> > Yes, removing FlushFileBuffers does solve the performance problem.
> >
> > I saw the most effect on large checkouts - you can 'hear' the difference
> > in disk activity as the files are written to disk on the client. The
> > server seems to do the same - I guess from flushing the last accessed
time
> > of the file.
> >
> > I can get some comparison timings if you like.
> >
> If it's going to make that much difference I'll push out another release
on
> Monday.  That'll give me enough time to make sure removing it doesn't
cause
> too many problems and pick up any feedback from this release over fixable
> bugs.
>
> Tony
>