 Samuel Thibault
|
2f6fe20825
Only reuse memchunks which are easy to throw away
|
8 years ago |
|
Samuel Thibault
|
72453f9f98
Disable forcing available memory by default, since this makes StarPU spuriously free data when prefetching fills the memory. Clean buffers are enough to be able to allocate data easily anyway.
|
8 years ago |
|
Samuel Thibault
|
33399e8f5a
Improve data reuse: before calling generic reclaiming, try to reuse potentially-in-use mc of the same interface. Avoids having generic reclaiming free data just to be able to reallocate exactly the same kind of data
|
8 years ago |
|
Samuel Thibault
|
65dc063a53
drop duplicate line
|
8 years ago |
 Nathalie Furmento
|
553f2b122e
doc: minor fixes
|
8 years ago |
|
Samuel Thibault
|
b1750358b4
Re-re-revert that pending trick
|
8 years ago |
|
Samuel Thibault
|
3ee0d40b90
fix comment
|
8 years ago |
|
Samuel Thibault
|
ef31b0b57f
Set diduse before calling the callback, for the case where the callback releases the data which was lazily unregistered
|
8 years ago |
|
Nathalie Furmento
|
0ab9a7a78c
doc: various fixes - add missing stuff
|
8 years ago |
 Corentin Salingue
|
08f4d332a3
fix weird behavior when drawing curves in ViTE
|
8 years ago |
|
Nathalie Furmento
|
66efac4d38
include/starpu_disk.h: make variable extern
|
8 years ago |
|
Samuel Thibault
|
bc0d878278
Expose the disk swap node number to the application
|
8 years ago |
|
Samuel Thibault
|
ffcd176304
Fix unit for STARPU_DISK_SWAP_SIZE: Bytes is really too small, use MB just like LIMIT*
|
8 years ago |
|
Samuel Thibault
|
91f985c664
Do not bother telling _starpu_post_data_request which node is handler for the request, it can determine it by itself from the request, and it is actually more right when the request is an invalidation, for which the handler is undetermined (-1)
|
8 years ago |
|
Samuel Thibault
|
3bb41b331a
Do not test the node kind if it is -1
|
8 years ago |
|
Samuel Thibault
|
8235eea5c7
Fix more disk cases: when just invalidating memory, we still need to note that it's the RAM which will run the write invalidation
|
8 years ago |
|
Samuel Thibault
|
2ddd26a88f
Fix requesting W access on disks: they do not have a worker to make them progress, so use the source node as being handler for the transfer
|
8 years ago |
|
Samuel Thibault
|
322115ff53
Fix setting diduse for a memory chunk: _starpu_memchunk_recently_used does not necessarily mean it was actually used, that could be a prefetch only. Make diduse marked from the tasks and from the data acquisitions instead
|
8 years ago |
|
Samuel Thibault
|
b0288c2d64
Show total gflop for each machine
|
8 years ago |
|
Samuel Thibault
|
b4012ec7ea
Avoid keeping all GPU communications in memory, process them while processing the trace
|
8 years ago |
|
Corentin Salingue
|
967b191cc3
Test return of malloc
|
8 years ago |
|
Corentin Salingue
|
2cfd0c3c44
Wait MPI sink workers finished before deleting node
|
8 years ago |
|
Samuel Thibault
|
2c9bbfd30f
Add functions to test for cached values
|
8 years ago |
|
Samuel Thibault
|
9e4166e8ed
Fix mpi cache flush documentation: one doesn't need to call it on strictly all nodes
|
8 years ago |
 Luka Stanisic
|
d96606dc31
traces: adding prefix for the scheduler name to comply with paje standard
|
8 years ago |
|
Nathalie Furmento
|
425c8442ab
ChangeLog: add missing stuff for 1.2.1
|
8 years ago |
|
Nathalie Furmento
|
10ddd00907
minor fixes for gcc warnings
|
8 years ago |
|
Nathalie Furmento
|
815ff2ac31
mpi/src/starpu_mpi.c: fix #19980, the code was not dead
|
8 years ago |
|
Nathalie Furmento
|
7f4969a2a6
mpi/src/starpu_mpi.c: remove dead code
|
8 years ago |
|
Nathalie Furmento
|
e33aad694a
merge branches/mpi_engine@18361: mpi: more code cleaning
|
8 years ago |
