[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Official Git mirror?

From: Eli Zaretskii
Subject: Re: Official Git mirror?
Date: Mon, 21 Feb 2011 23:15:05 +0200

> From: Óscar Fuentes <address@hidden>
> Date: Mon, 21 Feb 2011 21:57:30 +0100
> >> > The "smart" part is for sending less data, which is not going to win
> >> > for the initial checkout.
> >> 
> >> You said on the other post that cloning time is network-bound. So being
> >> smart and sending less data would be better.
> >
> > Not if "being smart" wastes CPU cycles on the server side and causes
> > it to use the available bandwidth less efficiently.  See the network
> > throughput figures above, reported by bzr on .bzr.log.
> Eli, I re-quote what you said:
> Oscar:
> >> >> It seems that "nosmart" is used for compensating for servers with busy
> >> >> CPUs
> Eli:
> >> > No, it's used to compensate for overly "smart" server when there's no
> >> > win in being smart, because you need to send everything anyway.
> You are contradicting yourself, because your data above is more evidence
> supporting my hypotheses of nosmart being a tricky effective with
> CPU-starved servers.

I see no contradiction.  The smart server uses more CPU on the server
side, and by that causes large gaps in sending data, while it "thinks"
what to send next.  When it needs to send hundreds of MBs, this adds
up to many minutes.  And in the initial checkout case, the invested
CPU time goes to waste anyway, because there's no way of being "smart"
when you need to send all the data downstream.

As an example, look at the "Finding revisions" stage, the first stage
of "bzr branch".  It takes less than a minute with "nosmart", and
around 20 min (!) with a "smart" server.  It tries to be smart about
which revisions to send, only to discover that it eventually needs to
send everything.

IOW, the "smart server" should be made aware that this is an initial
checkout into a fresh repository.  I guess this was not done yet
because for small projects the slowdown is barely visible.

> Besides, cloning from Savannah is almost twice as slow than from
> Launchpad for me. If it is because network distance considerations
> (pinging to it is 3 times slower than to Launchpad) or because the
> server is CPU and/or bandwidth starved, I don't know. Your timings
> comparing smart/nosmart seem to point to the CPU, though.

Launchpad is faster because its server is a newer bzr than that on
Savannah.  I'm not sure if the network bandwidth is also a factor
here, but it could be.

reply via email to

[Prev in Thread] Current Thread [Next in Thread]