Conversation
@kernellogger
@sj you might be interested in adapting DAMON parameters?
btw I wonder how broad range of parameters could be covered
@sj you might be interested in adapting DAMON parameters?
btw I wonder how broad range of parameters could be covered
2
0
0
@hyeyoo @kernellogger @sj @hyeyoo @kernellogger This strikes me though as a flaw in the kernel - if bpf can sit there in the background autotuning shouldn't that just be kernel code?
My concern with stuff like this is you end up with a mechanism that should be something sat in the kernel now sat outside of it.
I mean I do think tunables are sometimes used as a lazy option for something that really OUGHT to be autotuned but dev says 'meh leave it up to the user' and washes their hands of the whole thing.
Many tunables are very unclear to a user who doesn't know deep internals how to set and a very tiny % probably EVER get modified anywhere.
I wish we could just set fire to most of them but they become part of the uAPI so I guess we can't :)
My concern with stuff like this is you end up with a mechanism that should be something sat in the kernel now sat outside of it.
I mean I do think tunables are sometimes used as a lazy option for something that really OUGHT to be autotuned but dev says 'meh leave it up to the user' and washes their hands of the whole thing.
Many tunables are very unclear to a user who doesn't know deep internals how to set and a very tiny % probably EVER get modified anywhere.
I wish we could just set fire to most of them but they become part of the uAPI so I guess we can't :)
1
0
4
@ljs @kernellogger @sj
Ah, I added sj because IIRC he's planning implement an auto tuner for DAMON parameters - no one should be like "bpftune can tune everything and make more parameters"
it would have been horrible if i.e. active/inactive list balancing was implemented as a kernel parameter :P
Ah, I added sj because IIRC he's planning implement an auto tuner for DAMON parameters - no one should be like "bpftune can tune everything and make more parameters"
it would have been horrible if i.e. active/inactive list balancing was implemented as a kernel parameter :P
1
0
1
@hyeyoo @kernellogger Obviously very interesting topic. To my humble understanding, the point of this awesome project is the lightweight and fine-grained monitoring and adaptive tuning, which comes from the underlying use of BPF. That makes sense to me, as BPF is awesome. Nevertheless, it doesn't look like the panacea for every tunable knobs. Rather than that, it seems need per-case tuning implementation. I think it might be able to be useful for DAMON parameters, but tbh I'm not sure. Anyway, interesting work. Hope to have more chance to hear about that :)
0
0
3