Re: [HTCondor-users] zswap for jobs

[Emily Kooistra <a66@xxxxxxxxx>]

Hi All,

On 4/9/26 14:06, Oliver Freyermuth wrote:

> I second the "admin configurable" point raised by Emily ;-).

So Based on all input, I did some testing, and I think it just boils 
down to changing the behavior to the following:
- Configurable SWAP limit.
- Check if the kernel supports memory.zswap.writeback and disabling that.
- Just leave the zswap limit at max.

Now the "complicated" part, and not sure how people who have experience 
with it feel about this. You can configure the zswap pool size. Then you 
probably want a configurable way to express the total node memory. Or 
this includes the zswap pool by a factor. or you subtract it from the 
total node memory but give each job a boost factor (to some degree).

Altho then you also want to configure the limit when a job gets put on 
hold if it exceeds the memory limit (to some degree if possible).

This all does explode the option space a bit but good to think about it.

Altho it could be maybe even easier solved to have a classad to set 
cgroup options per slot. Given i currently don't see any way as a admin 
to set specific cgroup settings for Dynamic slots.

Emily