topic Re: data placement in Intel® Moderncode for Parallel Architectures

data placement

karimfath — Thu, 20 Nov 2008 22:15:45 GMT

hello

can i mesure the cost when a thread access data in l2 or ram to a do write or read operation.to evaluate difference between the two operation when data is in l2 and data is in ram in centrino duo system

thanks

Re: data placement

TimP — Fri, 21 Nov 2008 14:46:48 GMT

Quoting - karimfath

hello

can i mesure the cost when a thread access data in l2 or ram to a do write or read operation.to evaluate difference between the two operation when data is in l2 and data is in ram in centrino duo system

thanks

Have you considered to what extent PTU will help with your investigation?

http://software.intel.com/en-us/articles/intel-performance-tuning-utility-31-update-3

Re: data placement

Dmitry_Vyukov — Tue, 25 Nov 2008 09:56:58 GMT

Quoting - karimfath

You must make something like this:

- start 2 threads, bind them to different cores

- create following object:

struct X
{

char pad1 [128];

int data;

char pad2 [128];

};

X* g_x = new X;

- create semaphore for communication between threads

In order to measure cost when data is in L2$ and in foreign L1$:

1. First thread writes to g_x->data, and signals semaphore

2. Second thread wakes up and loads g_x->data

In order to measure cost when data is in RAM, you must flush data from cache with clflush instruction.

Re: data placement

DDd1 — Wed, 26 Nov 2008 23:52:53 GMT

Quoting - Dmitriy V'jukov

You must make something like this:

- start 2 threads, bind them to different cores

- create following object:

struct X
{

char pad1 [128];

int data;

char pad2 [128];

};

X* g_x = new X;

- create semaphore for communication between threads

In order to measure cost when data is in L2$ and in foreign L1$:

1. First thread writes to g_x->data, and signals semaphore

2. Second thread wakes up and loads g_x->data

In order to measure cost when data is in RAM, you must flush data from cache with clflush instruction.

Nice tip. Does anyone know of a small utility library that packs together small tips like this to do parallel performance measurements, and instrumentation in our own applications? Probably something they have laying around and can be of use to others ;)