Name: Anonymous 2016-04-18 2:46
http://www.ece.ubc.ca/~sasha/papers/eurosys16-final29.pdf
A decade of stovepipe additions to the Completely Fair Scheduler have resulted in a bloated, inefficient mess that does the opposite of what everyone thinks it's supposed to do. The result is that there are many edge cases where threads can wait for as long as seconds before they are assigned to a core.
The researchers were able to fix some of the scheduler bugs where there was a 23% performance improvement on a popular database with TPC or a shocking 137x performance improvement on HPC workloads. However, not at all of the scheduler issues can be easily addressed with simple patches but would require a more thorough redesign.
The researchers concluded, "more research must be directed towards implementing an efficient and reliable scheduler for multicore architectures!"
A decade of stovepipe additions to the Completely Fair Scheduler have resulted in a bloated, inefficient mess that does the opposite of what everyone thinks it's supposed to do. The result is that there are many edge cases where threads can wait for as long as seconds before they are assigned to a core.
The researchers were able to fix some of the scheduler bugs where there was a 23% performance improvement on a popular database with TPC or a shocking 137x performance improvement on HPC workloads. However, not at all of the scheduler issues can be easily addressed with simple patches but would require a more thorough redesign.
The researchers concluded, "more research must be directed towards implementing an efficient and reliable scheduler for multicore architectures!"