Re: Comparing Top 5 Timed Events and Cache Advisory

>>
>> But I can guarantee you that if CPU is, say, 70% and I/O wait 30%,
>> then the most you can gain by only adding memory is a portion of the I/
>> O wait, ie, a portion of 30%. It will never exceed that.
>
> I don't see how you can guarantee that. What if the 70% is due to cpu
> run queues from i/o asking for cpu to perform i/o,

that would not show as user mode CPU,that would be wait I/O or kernel CPU. I am assuming of course an OS that measures the correct quantities. Some, unfortunately, don't... I should have said the 70%CPU meant the user mode kind.

> certain OS patches have done something like that). As with every
> complex system, it depends, even a batch process is complex.

Exactly. But that complexity does not subvert CPU modes.

> Sometimes the answer is as simple as turning on async i/o 8-)

Very true. The numbers provided in the OP really don't give us much info in what regards the OS: they are mostly to do with the wait events and the cache advisory. That does not necessarily translate into "add more memory" at the OS level, a world of other things might be involved. More than likely, an inefficient process is at the core.

> I think we're all agreed that one needs to at least look at whether
> the work is optimal before throwing memory at it. In olden days,
> memory could help just about any system, but now even low end
> commodity machines may have enough.

Bingo. In general throwing more hardware only works well if all else is reasonably well sorted out. Throw more memory at an inefficient process and more than likely the result will be more memory inefficiently used! :)

I'm reminded of a recent series of events at work. After years of me pushing for partitioning to be extensively used in our DW, they finally started to do it. Performance improvement has been so large the folks in the DW team now are, if anything, over-partitioning!

Before, all I ever heard from them was: "we need a bigger box, this is too slow". Now, it's: "we're getting more than an order of magnitude performance improvement with LESS indexing! Why didn't you tell us indexes are slow?".

Love it! :) Received on Fri Jun 17 2011 - 19:45:31 CDT