 Marc Sergent
|
eed33d83d3
Merge from trunk @11911:11916
|
12 anni fa |
 Nathalie Furmento
|
c1b08ea04e
merge branches/starpu-1.1@11914: src/core/workers.c: create the pthread key earlier to make sure it is initialised when it is accessed for the first time
|
12 anni fa |
|
Marc Sergent
|
157dd97956
src/sched_policies/component_worker.c : Fix a false-positioned assert
|
12 anni fa |
|
Marc Sergent
|
913b6e242b
Several renamings for Hierarchical Schedulers to adapt the naming
|
12 anni fa |
|
Marc Sergent
|
638f30cba2
Merge from trunk @11789:11910
|
12 anni fa |
 Andra Hugo
|
1622267740
remove forgot printf in prev commit
|
12 anni fa |
|
Andra Hugo
|
4bd482e7b9
move hack out of the include files
|
12 anni fa |
|
Andra Hugo
|
e28f7852a3
hypervisor: steal work from other ctxs when the worker is idle
|
12 anni fa |
 Samuel Thibault
|
2fc5035219
Submit MPI data migration asynchronously, so it's all overlapped
|
12 anni fa |
|
Samuel Thibault
|
57720a2a82
Fix new features which actually appeared in 1.1
|
12 anni fa |
|
Samuel Thibault
|
c437b85f8b
Document how to perform MPI data migration
|
12 anni fa |
|
Samuel Thibault
|
c5bef53cf6
port r11793 from 1.1: Flush output at each line, to be able to check out progression
|
12 anni fa |
|
Samuel Thibault
|
96e6e58776
port r11793 from 1.1: Flush output at each line, to be able to check out progression
|
12 anni fa |
|
Samuel Thibault
|
804e6f8233
properly test for task termination
|
12 anni fa |
|
Samuel Thibault
|
e7d7ce19d4
typo
|
12 anni fa |
|
Marc Sergent
|
0026768555
Adding the handling of multiple fathers in hierarchical schedulers,
|
12 anni fa |
|
Marc Sergent
|
e998e008c6
Merge from trunk @11714:11788
|
12 anni fa |
|
Andra Hugo
|
2769aaa8cb
hypervisor: correct computation of idle
|
12 anni fa |
|
Samuel Thibault
|
4d81325ef7
fix missing initialization
|
12 anni fa |
|
Samuel Thibault
|
b8062a9fba
keep pushed back tasks in proper order
|
12 anni fa |
|
Andra Hugo
|
20906d8d77
hyp only: optimise resize (use idle and speed for the triggering policy)
|
12 anni fa |
|
Andra Hugo
|
6266721e89
do not take the mutex if the current worker is calling the code
|
12 anni fa |
|
Andra Hugo
|
39b06ba608
do not take changing_ctx_mutex when finishing executing the tasks (in decrement_nsubmitted_tasks), just check if the counter is 1, do what you have to do, and then decrement it (decrement the counter while verifying it wakes up the the thread calling wait_for_all)
|
12 anni fa |
|
Marc Sergent
|
fe685563e7
Adding several features which needs each other to compute correctly :
|
12 anni fa |
|
Samuel Thibault
|
b70f84e342
Properly find the task which optimizes computation time benefit
|
12 anni fa |
|
Samuel Thibault
|
a62e0dba32
better comment
|
12 anni fa |
|
Samuel Thibault
|
a06ea29938
Use separate expected ends, to avoid systematically preferring small tasks, possibly leading to starvation
|
12 anni fa |
|
Samuel Thibault
|
4f82d35ce9
Avoid keeping the lock held while scheduling tasks, to let scheduling happen in parallel. This means several processors might take 5 tasks each, thus emptying the fifo, but at least we don't have a huge contention point
|
12 anni fa |
|
Samuel Thibault
|
23b0ea64b9
Add a heft variant which draws tasks from its own fifo
|
12 anni fa |
|
Samuel Thibault
|
30c9f2732c
Plug perfmodel_select children normally from the scheduling strategy, instead of internally within perfmodel_select. This makes the room call go through the MCT node
|
12 anni fa |
