[Gllug] make -j problems

Nix nix at esperi.org.uk
Mon Jun 28 05:15:19 UTC 2010


On 27 Jun 2010, Richard Jones spake thusly:

> On Sat, Jun 26, 2010 at 10:24:20PM +0100, Nix wrote:
>>    Compression yes
>
> Assuming this affects scp as well (I think so) then I'll just note
> that on LANs this is not necessarily a win.

True. On my (private, trusted) LAN I actually go the whole hog (using
a pig-ugly hack, so the metaphor is accurate):

Host *.esperi.org.uk # and a big list
   Compression no
   TCPKeepAlive no
   ForwardX11 no
   ForwardX11Trusted yes
   SendEnv TERM DISPLAY X_ORIGINATING_HOST LANG LC_*

where the /etc/X11/xinitrc runs

export X_ORIGINATING_HOST="$(hostname)"

then every session-leading and login shell runs

export HOSTNAME="$(cat /etc/canonical-hostname)"
export DISPLAY

# ...

# If we are on a host remote to our initial X display but our DISPLAY
# does not reflect that, fix it up.
case "${DISPLAY:-}" in
    :*) [[ "$HOSTNAME" != "${X_ORIGINATING_HOST:-}" ]] && \
            DISPLAY="$X_ORIGINATING_HOST"$DISPLAY;;
esac

So I avoid the overhead of SSH X tunnelling and do X in the clear, while
getting the small performance benefit of unix-domain-socket- based
connections on the local host: if sshing to a different user, -X flips
on X tunnelling as usual.

I suspect this is hideously overdesigned and something with si:-based
authentication can do all of this on its own, but 1) si-based
authentication is completely undocumented and 2) given that, and that
this hack still works, I haven't bothered to rip it out yet...

>                                              I tested with/without
> compression extensively over LANs when writing the original P2V tool
> and compression of random disk images would slow things down a lot.

Oh yes, certainly.

I avoid this, in environments where I should use compression for most
transfers, by using different CNAMEs for bulk SSHes: because SSH
does no canonicalization at all when looking up the hostname in the
config file, this lets you use different bits of config for each.

> YMMV and the CPU/IO balance might have changed over the last two
> years

It's not even ignorable on Nehalems. Not quite, not if you have a
gigabit LAN. (Thousands of Kb/s versus tens of thousands.)
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug




More information about the GLLUG mailing list