How to scale hardware - monitor bus bandwidth

[darwinc]

The os and various agents can give us a clue on how busy a box is:

CPU % - is available
RAM usage - is available
Network usage - is available

But how do we see how much % available of memory bandwidth or disk i/o?

For example, the newer DL360 G6 have faster ram than the old but we don't really have a good idea how much do we need the newer hardware than the existing G5.

[David Claypool]

Well, it's not an easily expressed metric. Ideally, memory should be as fast as a processor (and the speed of a processor's memory bus should be equal to the processor's clock speed), but we left those days long ago. From a design perspective, you try to cope with that with cache, and hope that when a CPU retrieves a piece of data that it can be found in Level 1, which operates as close to the processor's speed as possible. If it isn't, then it has to re-fill from Level 2 (some processors even have a L3 cache) until finally from main memory.

With faster main memory, the faster a cache fill request can happen. Statistically speaking, with more cores per processor, the likelihood of a cache miss goes up, so the more main memory fill operations are needed.

[juan quesada]

google

+++ performance counters site:*.microsoft.com
where +++ is the subsystem, role or application you want to know more about.

example:

memory performance counterssite :*.microsoft.com

JQ

[rick jones]

Being able to see the bus utilization depends on whether or not there are HW performance counters for it. You will need to find tools that display Intel/AMD HW performance counters and then run those.

100% CPU util may not mean bus saturation, but bus saturation will probably express itself as 100% CPU util.

[darwinc]

I will check on the google side. I suspect that the hw stats (via windows os) will not be available if it is running in vmware enterprise.

In the data center:
An old proliant 1500 (pentium pro) shows about 20% cpu utilization when it is zipping some files.

However, a current workstation can show similar cpu utilization, while crunching 10x the above workload. (and get it done faster)

Even, if I remove the diskchannel from the equation, I can still get similar cpu % on various systems. Basically, giving folks no clue on where is the slowest part.

so, anyways, back to issue at hand.

Faster is always better (grin), but, is it not possible on HP servers to see how busy the various subsystems are?

So, then certain upgrades can be justified to folks that would rather just look at the CPU % only !

[rick jones]

Indeed, I think the virutalization of HW performance counters is still a research topic :) You would have to run the HW counter tool in the hypervisor rather than a guest OS.