Re: [Monotone-devel] performance profiling

[Nathan Myers]

On Sun, Apr 17, 2005 at 06:25:24PM +0200, Sebastian Spaeth wrote:
> I tried oprofile to get *some* idea where CPU cycles go, so I pulled a
> fresh monotone database. 
> 
> monotone: [bytes in: 22.6M] [bytes out: 624.4k] [certs in: 6965] [revs
> in: 2441]
> real    32m17.546s
> user    27m58.065s
> sys     0m11.042s
> 
> samples  %        symbol name
> 8063048  31.5207  CryptoPP::Deflator::LongestMatch(unsigned int&) const
> 2182601   8.5324  CryptoPP::Deflator::ProcessBuffer()
> 1401468   5.4787  CryptoPP::LowFirstBitWriter::PutBits(unsigned long,
> unsigned int)
> 1190379   4.6535  CryptoPP::Deflator::EncodeBlock(bool, unsigned int)
> 988320    3.8636  CryptoPP::BaseN_Encoder::Put2(unsigned char const*,
> unsigned int, int, bool)
> 586569    2.2931  CryptoPP::Deflator::FillWindow(unsigned char const*,
> unsigned int)
> 
> Interesting that we call LongestMatch 8mio times for 22Mb received data.
> sounds a lot. Does anybody know how oprofile can tell me how much time
> was spend in these functions?

It makes me wonder if the source repo might not just send its already-
compressed images, so the client doesn't need to compress them all over
again.  (The certs are on the compressed versions anyway, right?)
Or are they sent compressed, but getting unpacked and repacked locally?

Nathan Myers
address@hidden