Re: [m5-users] Multithreading in m5

Korey Sewell Mon, 09 Oct 2006 16:54:48 -0700

Hi Rajdeep,

It seems like you want to study some type of thread-scheduling policy across multiple CPUs. I think M5 can handle what you want given that you are able to modify some code to hook in your thread-scheduling policies.

It seems to me that you would have to modify the code in two places:

(1) The system call table in process.cc - These are defined for each architecture-OS pair in the src/arch directory (e.g. src/arch/alpha/linux or src/arch/alpha/tru64/ or etc. etc.)

(2) The system call declaration & definition in src/sim/syscall_emul.hh & src/sim/syscall_emul.cc

Basically, the table in the aforementioned 'process.cc' is a map of the names of system calls and function pointers to their appropriate functions definitions (usually in syscall_emul.hh/cc) In that table you will find system calls like "fork", "open", "write", etc.

Some of the calls are unimplemented (which map to say 'unimplementedFunc' function pointer) while others are implemented.

I didnt see a function defined for the 'fork' in the tables but I think it shouldnt be too hard if you look at the definition for 'nxm_thread_create' as an example. That syscall is for the alpha-tru64 combo, where the system call is mapped in src/alpha/tru64/process.cc (#32), and it is defined in src/kern/tru64/tru64.hh.

If you are just doing a syscall-emulation mode simulation I think the bottom of the definition would be most useful, where the code right now looks like:

// Find a free simulator execution context.
for (int i = 0; i < process->numCpus(); ++i) {
ExecContext *xc = process->execContexts[i];

                if (xc->status() == ExecContext::Suspended) {
                    // inactive context... grab it
                    init_exec_context(xc, attrp, uniq_val);

...

What's tricky here is that process->numCpus() is a bit of a misnomer since it really is giving you the number of thread contexts available in the system. In a multicore simulation with one thread per CPU then you are fine. But in a multicore simulation where there are multiple threads per CPU you would also need to check the 'ExecContext' "cpu_id" for whatever scheduling purposes are appropriate.

So I guess your 'fork' implementation might have a similar loop with some of the aforementioned changes.

If you want to specify what workloads are assigned to specific cores at the beginning of simulation then that would be something to handle in your M5 configuration script (and I didnt have to write this long email!).

Anyways, I hope what I wrote above helps....

On 10/7/06, Rajdeep Bhowmik <[EMAIL PROTECTED]> wrote:

Hi,

I want to use the m5 simulator to learn how multi-threaded matrix
multiplication code will work on a multi-core architecture. Is it
possible for me to specify how the various threads of this code should
be mapped to the different cores? Or, does m5 have any special library
calls to spawn off threads to different cores (eg: equivalent of a
fork() call).

regards-

Rajdeep

__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
http://mail.yahoo.com

_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

--
--
Korey LaMar Sewell
University of Michigan
Computer Science & Engineering
Ph.D Student
------
``Experience is not what happens to you. It is what you do with what happens
to you.''

_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Re: [m5-users] Multithreading in m5

Reply via email to