LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: Hitting the real Limits

To: ratz <ratz@xxxxxx>
Subject: Re: Hitting the real Limits
Cc: Thomas Proell <Thomas.Proell@xxxxxxxxxx>, Joseph Mack <mack.joseph@xxxxxxx>, <edward@xxxxxxxxxxxxxxxxxx>, "lvs-users@xxxxxxxxxxxxxxxxxxxxxx" <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
From: Julian Anastasov <ja@xxxxxx>
Date: Tue, 14 Nov 2000 01:11:20 +0000 (GMT)
        Hello,

On Mon, 13 Nov 2000, ratz wrote:

> Hi,
>
> > Or tunneling. I wrote hin already that I can "satisfy" up to 70000
> > clients without delay on a Pentium1 (simulated with testLVS). And there,
> > I think the 10MBps network is the problem, not the processor, because
> > "top" shows a load of 0.14  :-)

        Yes, 10Mbps can be a bottleneck even for slow CPUs.

> Hmmm, does top really show the kernel space usage of a process/task?
> How about kernel threads? I guess this is the value of system, isn't it?
> I mean, you can't see how much 'load' the machine has for example whilst
> packetfiltering.
>
> Could someone please elaborate it a little bit more for me please (Julian)?

        The load average is calculated from the number of active processes:
TASK_RUNNING, TASK_UNINTERRUPTIBLE or TASK_SWAPPING. It seems the
processes in the last state (TASK_SWAPPING) are not included in Linux 2.4.

        The routers can experience load average 0.00 even on full links,
for example, if there are no running proccesses.

        But you can start many running processes and to try to measure the
real CPU load from the network system (not sure how the load average
will increase for looping processes on different load from the networking).
But you can start a standalone bogomips test or similar tests that
measure the CPU cycles for specific time interval.

> Best regards,
> Roberto Nibali, ratz


Regards

--
Julian Anastasov <ja@xxxxxx>



<Prev in Thread] Current Thread [Next in Thread>