rhea: scale less
It's been commonly hapenning that we spin up many machines but can't keep them occupied due to bottlenecks in the central machine (probably it's mostly the compression of copying-results step) So let's scale less aggressively and thus waste less.
Vladimír Čunát committed
Jan 20, 2024 at 09:31 UTC
3638c927dd5bbffe8392336cd37444f1b6654730
1 file changed
+7
-7
delft/hydra-scaler.nix
+7
-7
@@ -15,9 +15,9 @@
15
categories = {
16
aarch64-linux = {
17
bigparallel = {
18
- divisor = 5;
18
+ divisor = 20;
19
minimum = 1;
20
- maximum = 5;
20
+ maximum = 4;
21
plans = [
22
{
23
bid = 2.0;
@@ -27,9 +27,9 @@
27
];
28
};
29
small = {
30
- divisor = 2000;
31
- minimum = 1;
32
- maximum = 5;
30
+ divisor = 4000;
31
+ minimum = 0; # we have one static 80-core in addition
32
+ maximum = 3;
33
plans = [
34
{
35
bid = 2.0;
@@ -41,7 +41,7 @@
41
};
42
x86_64-linux = rec {
43
bigparallel = {
44
- divisor = 5;
44
+ divisor = 10;
45
minimum = 1;
46
maximum = 5;
47
plans = [
@@ -60,7 +60,7 @@
60
small = {
61
divisor = 2000;
62
minimum = 1;
63
- maximum = 5;
63
+ maximum = 4;
64
plans = [
65
{
66
bid = 2.0;