Is my calculation correct for L2 miss rate using VTune?

[mfcking]

Hi,

I want to compute how many L2 cache misses each packet has during IPv4packet forwarding in Linux network stack? I use P4 Dual Xeon and Linux 2.6 kernel with VTune 3.0.

I use VTune to get a L2 load cache miss retired events, divide this by the total instruction retired to get a L2 cache miss rate per instruction, then multiply this with the number of instructions per packets to get a L2 miss per packets.

However, the L2 cache miss rate per instruction is 3.4% and each packet needs about 2000 instructions, thus ittakes roughly has60~70 L2 cache miss when processing each packet. I'm not sure this calculation is correct or not?

Thanks a lot,

Liang

Message Edited by mfcking@yahoo.com on 08-29-2005 12:04 PM

[TimP]

L2 Load miss rate per instruction should already be present among the pre-programmed ratios available for collection. It's calculated from 2nd level cache load misses retired, and should save you having to worry about different sampling rates. In the version I have installed, it's among the "Primary tuning parameter" ratios.

[sharshera]

Where exactly did you get those preprogrammed ratios?

Which build do you have for Vtune? Is that a Windows Vtune or a Linux one?

[TimP]

Yes, it's the Windows analyzer, running rdc on linux.

[sharshera]

Thanks very much for the tip.
I have a linux 3.0 VTune and it doesn't have all of these features of the windows version.

So, to have this setup, you would have to buy Vtune for Windows and then the rdc for linux or what?

[TimP]

Vtune for Windows provides remote collectors for Windows and linux, and access to premier.intel.com downloads of new rdc. Maybe not the place to mention it, but 64-bit linux rdc no longer works for me with Windows 2000; it seems most likely to work with 32-bit XP (including SP2). Don't know about 64-bit Windows, which I would prefer if 2K no longer works.