search

rvanheest / IN4392_Cloud_Computing_Labs

0 stars 0 forks source link

Experiment suggestion #23

Open rvanheest opened 10 years ago

rvanheest commented 10 years ago

Can we do an experiment comparing the QueueLengthScheduler with the BlockingQueueLengthScheduler?

As I'm writing in the report about their differences, I write that the workers can focus more on processing the current image rather than being interrupted all the time by receiving new tasks when the tasks stay as long as possible in the head node.

Froussios commented 10 years ago

Yes we can test that, but I am certain that there is no such advantage. Unless I missunderstand what you're trying to say. Why does the blocking variant interrupt less? The total amount of work in a worker (incuding communication) is independant of the scheduling policy. What changes is how you make sure everyone is busy at all times.

rvanheest commented 10 years ago

I know, but when you are a single core worker and you receive new tasks all the time, you can't spend that time on processing the current task. That's what I tried to say.

Froussios commented 10 years ago

Interrupting is done once per job, on average. The average processing time should be unaffected (because switching threads is extremely fast compared to our jobs), but there might be a slight change in the variance. Is that what you want to measure?