From: Edwin Vollebregt Newsgroups: comp.parallel.pvm Subject: What does PvmPollTime do ? Date: Wed, 16 Jun 1999 07:13:54 +0200 Organization: Delft University of Technology Message-Id: <37673292.785A1EFF@pa.twi.tudelft.nl> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.parallel.pvm:8490 Hello, In our application we use multiple computing processes per processor. This works fine as long as each computing process goes to sleep when it enters a call to pvmfrecv or pvmftrecv. However, on different multi-processor platforms (SUNMP, HPPAMP, SGIMP), using PVM 3.3.11, we noticed that the call to pvmftrecv does not cause a process to sleep. We have seen the same happen on different machines with pvmfrecv, but did not note down the circumstances. On the HPPAMP architecture the same happens with PVM 3.4.0 if PvmPollTime=1000000 (default value). If PvmPollTime=0 (and PvmPollType= PvmPollSleep), then pvmfrecv does not cost significant cpu-time, but pvmftrecv still consumes cpu-time. Can anyone tell me more about the mechanisms used in pvmfrecv and pvmftrecv on SMP-systems ? Especially I would like to know why PvmPollTime does not affect the latter, and am trying to reduce the cpu-time needed for this case. If waiting for messages costs cpu-time then this causes a dramatic reduction of the parallel performance of our application. This is because the cpu-time that is consumed by a process waiting for a message delays the calculations in other processes, and thus also delays the arrival of the message. For instance if there is a token that is sent back and forth between two processes that each work on the token for about a second, then one step takes two seconds if waiting does not cost cpu-time, but it takes four seconds if waiting costs full cpu-time. This is because the processor that holds the token gets only half of the cpu-capacity. Any help is greatly appreciated. Edwin Vollebregt Delft University of Technology / VORtech Computing edwin@pa.twi.tudelft.nl