Software Prefetching using Machine learning

[Rik van Riel]

On Thu, 2019-10-10 at 11:10 +0900, Irfan Ullah (울라 이르판) wrote:
> Thank you  for the responses.
> It is not my assignment. Actually, we have initiated a project based
> on the idea "incorporating artificial intelligence in the linux
> kernel to increase its performance". 

That is an interesting idea, but it is worth keeping in
mind that most "kernel performance" is not about how well
the kernel does something, but about how quickly the kernel
lets userland programs do things.

For example, I have done a lot of work on the scheduler
code recently, and the most common thing I see (whenever
I try something wrong) is for total CPU use in the system
to go up. However, the increase in CPU time used is often
in user space, and not in the kernel.

For example, trying out a bad idea might result in 0.2%
more CPU time spent in the kernel, but 2% more CPU time
spent in user space, due to seeing a larger number of
cache misses from user space programs.

When things are done right, enabling a certain kernel
feature might result in some overhead (say 0.3%), which
is almost entirely caused by cache line misses, resulting
in a total system overhead of twice that (0.6%), because
every cache miss in my code results in something else
getting evicted from the cache, and somebody else incurring
a cache miss as well.

In short, if you want to use machine learning to improve
kernel performance, you need to figure out a way to do
that without causing cache misses in any really hot path.

That puts a serious constraint on what you can do.

-- 
All Rights Reversed.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: This is a digitally signed message part
URL: <http://lists.kernelnewbies.org/pipermail/kernelnewbies/attachments/20191010/c0143c32/attachment.sig>