 Samuel Thibault
|
6adb01dd56
port r11392 from 1.1: Do not hold the handle lock while allocating an interface
|
11 年之前 |
|
Samuel Thibault
|
245bdafca2
port r11388 from 1.1: Allocate RAM memory immediately, without using requests
|
11 年之前 |
|
Samuel Thibault
|
e1501b5592
port r11308 from 1.1: record allocation size in trace
|
11 年之前 |
|
Samuel Thibault
|
f850d2f26a
port r11305 from 1.1: the handle interfaces are already freed at this point, use the footprint pre-computed in the mc
|
11 年之前 |
|
Samuel Thibault
|
c4dc808530
port r11299 from 1.1: Use a hash table to speed up allocation cache reuse by hashing through the data footprint
|
11 年之前 |
|
Samuel Thibault
|
01aa31bde8
port r11297 from 1.1: Avoid looking through all data for not_important data if the application did not set any data as not_important.
|
11 年之前 |
|
Samuel Thibault
|
0bb629fed8
Avoid reverting the ordering of cache items
|
11 年之前 |
|
Samuel Thibault
|
12ca1d21d3
print the OOM node and how much we try to flush. Also print only once
|
12 年之前 |
 Nathalie Furmento
|
d339c20e4a
src/datawizard/memalloc.c: print message when trying to reclaim memory. It will not be printed if environment variable STARPU_SILENT is set
|
12 年之前 |
|
Samuel Thibault
|
065614f6b0
port r11042 from 1.1: Simplify LRU mechanism: now that the mc_rwlock is taken after data header lock, we can simply modify the list, instead of pushing the LRU order to a temporary list, which has a huge worst-case complexity (number of accesses * number of data)
|
12 年之前 |
|
Samuel Thibault
|
a9acc0f302
target can be -1...
|
12 年之前 |
|
Samuel Thibault
|
6d7b9fc372
Handle case when we have not found a target
|
12 年之前 |
|
Samuel Thibault
|
9c038ac2be
replace starpu_get_bandwidth/latency_CUDA/RAM_RAM/CUDA with calls to starpu_transfer_*
|
12 年之前 |
|
Nathalie Furmento
|
1bdf2585ae
merge trunk
|
12 年之前 |
 Marc Sergent
|
9783f21238
Fix of the livelock discovered by Marc Sergent's internship, by adding a maximum trylock threshold before a blocking lock
|
12 年之前 |
|
Marc Sergent
|
e0b771e409
FxT lock traces support added to StarPU. Use --enable-fxt-lock option of configure to activate it
|
12 年之前 |
 Corentin Salingue
|
e75bbec0cf
merge trunk
|
12 年之前 |
|
Samuel Thibault
|
582fe879e4
comments for the LRU algorithm
|
12 年之前 |
|
Corentin Salingue
|
210ed2e470
add disk flag + add flag in reclaim + add test disk_computation
|
12 年之前 |
|
Samuel Thibault
|
a063f6dd2b
Use STARPU_MAIN_RAM, not 0
|
12 年之前 |
|
Corentin Salingue
|
3d450f4b5e
correct bug target can be the reclaim node
|
12 年之前 |
|
Corentin Salingue
|
8281604526
merge trunk
|
12 年之前 |
|
Samuel Thibault
|
f5376f57de
Fix missing unlock when a data tree could be locked, but the data couldn't be freed
|
12 年之前 |
|
Corentin Salingue
|
209fb7bf53
merge trunk
|
12 年之前 |
|
Samuel Thibault
|
0bb47b1bab
Rework locking the mc_list: we must not release mc_rwlock while freeing a memchunk, otherwise some other part of starpu might be removing the very chunk we are trying to free. Keep mc_rwlock locked, and when one needs to lock the data header, use a trylock for the case when some other part of starpu has already locked the data and will lock mc_rwlock too
|
12 年之前 |
|
Samuel Thibault
|
cbd7a8bace
Add missing lock
|
12 年之前 |
|
Corentin Salingue
|
1705223a6c
forget case home_node == disk -> push on ram before + move target choice strategy into a function
|
12 年之前 |
|
Corentin Salingue
|
21c26a7564
clean function to choose target for reclaiming
|
12 年之前 |
|
Corentin Salingue
|
01181bc7ab
if data is allocated, we don't check space on the node
|
12 年之前 |
|
Corentin Salingue
|
4519397359
fail merge
|
12 年之前 |
