(FWD)Linus on fast UDB networking.

[gaf at zk3.dec.com (Jerry Feldman)]

Forwarded from axp-list at redhat.com
From: Linus Torvalds <torvalds at cs.Helsinki.FI>
On Tue, 15 Oct 1996, Darryl Wagoner wrote:
> 
> When I decided to buy a UDB I felt stuck.  I could use Linux and
> have poor networking or I could use NetBSD/OpenBSD and be on the
> bleeding edge.  I opted to Linux as I don't like to bleed and
> don't have a lot of time if things go wrong.  I am not even sure
> that OpenBSD would even work.  

The days of poor Linux networking are _long_ gone..

> I did a quick read check and I got about 200k bytes per second
> over NFS.  Not great but not unusable either.
> 
> This morning added a 1.7 gig drive to the UDB and went to copy
> the Red Hat axp tree over to UDB via NFS.  I thought my Sparc II
> with a Weitek chip was going to die.  The Sparc couldn't keep up
> with the UDB.  This would completely unexpected!  
> 
> Way to go Alpha Linux team! You have done yourself proud!!!! 
> 
> Why is the Intel networking so bloody slow?

It isn't. Not any more.

Essentially, with 2.0.x kernels (or 2.1.x kernels) you can more-or-less count
on Linux being faster at networking than just about _anybody_ else.  Trust
me, we get some _very_ studly numbers, and it's not only on alphas.  We are
beating Solaris on the same hardware on sparcs, and we're beating the free
BSD's on intel. 

There is one thing that isn't very fast: NFS writing. That will get fixed
too, but in the meantime you should know about this, and try to avoid
depending on large writes over NFS (actually, let's rephrase that as "large
number of small writes", because that's really where the problem is). But
reading over NFS is certainly not slow, and UDP/TCP are up there with the
absolute best too. 

It seems your newfound speed is due to a faster harddisk. That can make a
rather huge difference, especially with program loading (and program loading
is one of those things you notice when it's a question of "snappyness" of a
machine). It may be that the original 340M drive in the UDB isn't all 
that fast.

(It may also be that you upgraded kernels: depending on what release you
originally installed with, your original kernel may have been old enough that
the NFS read cache wasn't in place. That certainly makes a difference,
especially if you have RAM enough to see some caching). 

		Linus

- --
To unsubscribe: mail -s unsubscribe axp-list-request at redhat.com < /dev/null


------- End of Forwarded Message