Andra Hugo
|
e28f7852a3
hypervisor: steal work from other ctxs when the worker is idle
|
11 years ago |
Andra Hugo
|
2769aaa8cb
hypervisor: correct computation of idle
|
11 years ago |
Andra Hugo
|
20906d8d77
hyp only: optimise resize (use idle and speed for the triggering policy)
|
11 years ago |
Andra Hugo
|
12dec4c5cc
hypervisor optimisation (avoid unnecessary notifications/locks/trylocks) + add the hyp the traces
|
11 years ago |
Andra Hugo
|
9bb7bdb07f
Increment/decrement submitted/ready tasks per context (instead of doing it per context & globally)
|
11 years ago |
Andra Hugo
|
5a22d83b5c
hypervisor: feft_lp: compute max by using the execution of the last sample instead of the idle time
|
11 years ago |
Andra Hugo
|
c4873d4813
hypervisor lp: max is computed using the time of executing tasks instead of idle time
|
11 years ago |
Andra Hugo
|
994728b26b
hypervisor lp: max computation for the contexts sharing resources ( do not consider shared workers as full workers)
|
11 years ago |
Andra Hugo
|
1289532233
whenever there are a few resources the lp cannot assign, all contexts share these resources
|
11 years ago |
Andra Hugo
|
082dc8501c
hypervisor: adapt lp solution when max has an important impact on the solution (exec lp twice for now, next co share resources)
|
11 years ago |
Andra Hugo
|
52459c8508
*eager: wake up workers when we add them to a context
|
11 years ago |
Andra Hugo
|
646c61a828
improve idle computations in order to better compute max_workers for the resizing lp
|
11 years ago |
Andra Hugo
|
9502f53f25
introduce min max as a constraint for the linear program
|
11 years ago |
Andra Hugo
|
9f20f53ce9
fix size_ctxs (do not consider ready tasks for size, only for resize), speed is computed on fixed value of the time
|
11 years ago |
Andra Hugo
|
f108a67609
add a prolog callback to a task (+ example)
|
11 years ago |
Andra Hugo
|
39cfbbcb7f
consider the flops of ready tasks when resizing and compute the speed for a sample of time too
|
11 years ago |
Andra Hugo
|
1c1a6aaddf
* add function to check if the workerid returned by the func starpu_worker_get_id belongs to a context,
|
11 years ago |
Andra Hugo
|
7e15608069
copy uthash.h into the hyp sources and remove any interferance of the hyp with the sources of starpu
|
11 years ago |
Andra Hugo
|
93ed656614
add missing hyp functions to the doc + some code cleaning
|
11 years ago |
Andra Hugo
|
7652adb53c
help for the hypervisor policies
|
11 years ago |
Andra Hugo
|
4bd4f6b009
bug fix for the prev commit
|
11 years ago |
Andra Hugo
|
51f869b6a1
reset flops when adding and removing workers too (not only moving)
|
11 years ago |
Andra Hugo
|
092e72d872
add some doc for the hyp and change name debit-> throughput
|
11 years ago |
Andra Hugo
|
d010c07b4a
add func to update the elapsed flops
|
11 years ago |
Andra Hugo
|
9d2ba463c5
remove some warnings
|
11 years ago |
Andra Hugo
|
109eac0300
remove contraint for tmax in lp
|
11 years ago |
Andra Hugo
|
abebb0c602
cleaning ispeed_lp code + free hwloc_worker_set when deleting the context
|
11 years ago |
Andra Hugo
|
50dfbc0ca5
change number of flops of a ctx: difference of flops instead of total flops
|
11 years ago |
Andra Hugo
|
0a164e277c
change dynamically the total number of flops of a context
|
11 years ago |
Andra Hugo
|
6e52308e70
Indicate whether a worker shares tasks lists with other workers (if yes if we remove it from the context we should do it right away, the other workers are in charge of the remaining tasks)
|
11 years ago |