Hi all!
Daily questions :-)
As I am tracking down the thread creation process in splash2(FFT) with SE
mode, I found that in tru64.hh, the system call stack_createFunc is executed
twice before the system call nxm_thread_createFunc is executed. I don't know
why pthread is implemented in this way since I thought stack_createFuns
needs only be called once. Do you by any chance have any clues?
What I want to do is to add a load balancer/scheduler to the system.
Instead of using pthread and monitoring threads with system calls, I want to
add new instructions and monitor threads with the scheduler, bypassing
related OS level system calls. I also assume that all the threads take a
single void* as their argument. On thread creation, the mainthread executes
new instructions to store pc and the void* in a packet, and then another
instruction sends the packet to the scheduler and signals the scheduler to
create a new thread.
My question is, what steps are there to create a thread without system
calls? My assumptions are:
0. pool some unallocated thread contexts on all the cpu/cores at the
beginning of the simulation
1. one thread creation, call stack_createFunc once
2. push in the void* as arguments to the stack
3. follow what nxm_thread_createFunc does and initialize the attrp,
call init_thread_context and activate a thread context in a cpu/core.
Is this the right thing to do or it's much more complex than what I see?
Btw, some questions regarding code in tru64.hh::nxm_thread_createFunc: What
is RAD and VP slots? Can you point me to any background references?
Thanks and have a nice weekend!
Jiayuan
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
