Re: Follow-up: CPU poll threads ...

On Wed, 28 Oct 1998, Art S. Kagel wrote: > Schaeflein, Paul wrote: > > > > Would you please describe the difference between NET VPs and CPU VPs? And > > why is one better than the other? > [Original reply SNIPPED] > > OK. Basically a NET VP (shown as class tli, soc, ipx, shm, str in the > onstat -g glo report) is a Virtual Processor whose only job is to poll> the various connections for new sessions and additional queries on > existing sessions and pass the work off to CPU VPs as needed. The CPU > VPs are doing actual query work for various sessions and in addition > can handle exactly one poll thread to do listening itself. Basically > it is best for networked connection types to have listeners in NET VPs > and shared memory connection listeners in CPU VPs. Shared memory > listeners in NET VPs will poll incessantly burning CPU time. Network > listener threads in CPU VPs use more system call overhead. The > difference has to do with the fact that the NET VPs have nothing better > to do than monitor the connections and so can block waiting for network > activity but they cannot block on shared memory operations and so have > to poll in a loop. Inversely the CPU VPs cannot block waiting for > network activity, they have better things to do with their time, and so > have to poll network connections each time they have completed some > user thread operation. Meanwhile the very fact that they are busy most > of the time doing real work means that they can only poll shared memory > connections occasionally which reduces the polling overhead for shared > memory connections versus NET VPs. > > Also it is important that ALL CPU VPs run listeners for the primary > shared memory connection not just some. Since the shared memory > polling only takes place when the VP is between tasks this improves > service times. In addition if only some CPU VPs are polling shared > memory, and they do so only when they are not busy, then the polling > VP that accepts the task will schedule itself to work on it since it is > free and the other CPU VPs will rarely get any tasks assigned to them > which keeps the polling VPs even busier and reduces service times even > further. > > Hope that's enough. If you need more check TFM. I would like to know *where* in TFM this info is discussed in this much detail. :) Steve Berg > > Art S. Kagel >