[geeks] Linux disparity

[Joshua D Boyd]

On Wed, Dec 18, 2002 at 12:39:00PM +0000, Mike Meredith wrote:
> On Wednesday 18 December 2002 8:01 am, Joshua D Boyd wrote:
> > Hey, on my linux machine, CPU meters register barely anything, yet I
> > have a very high system load (over 2).  Anyone know why this might
> > be? 
> 
> Do you mean the output of 'uptime' ? That's the average number of 
> processes waiting in the run queue. And '2' is hardly very high ... 
> I've seen Linux servers recover gracefully from states with load 
> averages in the 50's.

I thought it was the average length of time a process waited in the
queue.  I guess average number would make more sense.

Anyway, a load average of 2.0 - 4.0 or even 5.0 is quite comfortable on
SPARC hardware in my experience.  However, this linux/x86 machine does
not feel comfortable.  It is sluggish when the uptime number gets
there, and I'm trying to figure out why.  I think it is an issue with
networking and/or NFS specifically.

> 'top' is a far better way of analysing what your system is doing at any 
> one time. In particular look at the "system" percentage as that shows 
> the amount of work being done by the kernel rather than user processes 
> ... which might be relevant here.

I'll check that out.
 
> > I'm doing a fair amount of network copying in the background, but
> > nothing heavy over an SSH link.  Is it perhaps being caused by a 3COM
> > PCI ethernet card?  I seem to remeber people dissing them here.
> 
> I don't remember it coming up here, but nobody I work with would stick a 
> 3com network card into a server because of the load they are rumoured 
> to put onto the CPU.

Well, at the time I aquired the card, I was under the impression that
3com ethernet adapters were hot shit, and everything else was the cheap
replacement.  I just came up with a cheap netgear fast E card while
stripping a P166 (the card was put in that P166 in 2001 though, so it
ain't that old).  I don't know the chipset, but is probably pretty low
end for netgear or Verizon won't be including it for free to DSL
customers.  Is Netgear supposed to be better.
 
> The most important metric of performance is does the box feel slow ? If 
> it doesn't then it isn't worth trying to chase down what is causing the 
> problem (except out of curiosity).

Yes, it does feel quite slow.  Sure, doing a lot over the network isn't
as fast as local disk, but it is far slower than even that.  I'm just
trying to find a metric to measure this and trace it back to the cause.
Since it is caused by using NFS and/or FTP between my main linux machine
and linux file server.  I tried decreasing the packet size, and that
helped make things more stable and faster, but I still want to know why
I often get better performance over DSL than I am between these two machines.
 
> BTW: Has anyone got any really good references for performance analysis 

Yes please.
-- 
Joshua D. Boyd