Reduce ITP with number of active tests per user #1747
Conversation
given some excessive number of active tests by some users, this patch reduces itp based on the number of active tests. The formula used is `itp *= 36.0 / (36.0 + count * count)` This reduces slowly the itp for a number of tests smaller than 6, and rapidly afterwards. For 3 active tests, the itp is 80%, for 6 active tests, the itp is 50%, for 15 active tests it is down to 13%. For a maximum share of fishtest... have 6 tests in queue.
|
DEV updated. |
|
EDIT_000: updated values and screenshot
|
|
The precise numbers might need some time to establish, i.e. once the itp has been recomputed. With 9 tests, it should be around 30% |
|
which it is now: |
|
I updated the previous post, now the ITPs seem to be correct. |
|
I have a question, can we regard tests already set to lower throughput such as 50% as something like 0.5 tests |
|
no, not easily. To transfer throughput, one should be able to pause the others (i.e. use prio -1) |
|
I think this is working as intended |
|
PROD updated, thank you @vondele :) ps: nitpicking, 36.0 was mandatory in python2 |
This distributes the number of cores more fairly between users, and allows for an effective way to transfer/split TP between runs. (See official-stockfish#1747 (comment))
given some excessive number of active tests by some users, this patch reduces itp based on the number of active tests.
The formula used is
itp *= 36.0 / (36.0 + count * count)This reduces slowly the itp for a number of tests smaller than 6, and rapidly afterwards. For 3 active tests, the itp is 80%, for 6 active tests, the itp is 50%, for 15 active tests it is down to 13%.
For a maximum share of fishtest... have 6 tests in queue.