kirk@wolfen.cc.uow.oz (Kirk Barrett) (06/01/90)
G'Day, has anyone had any experience with vmtune in the area of tuning a Symmetry series to page instead of swap. We have am S27 with 8M of memory and as soon as 2 or more people start running a certain statistics package which uses a fair bit of memory, the system starts swapping, the load average hits the roof (upto 21) and response time etc. drop considerably. It has been suggested that I could tune the system to page instead of swap with vmtune but that this may be tricky. Any Ideas or experiences ??? Thanks in advance Kirk Barrett ACSnet: kirk@wolfen.cs.uow.edu.au Consultant Programmer UUCP: ...!munnari!wolfen.cs.uow.edu.au!kirk Computing Services ARPA: kirk%wolfen.cs.uow.edu.au@uunet.UU.NET University of Wollongong Phone: +61 42 270121 N.S.W. 2500 Australia Fax : +61 42 297768
peter@sigma21.oz (Peter Farrell) (06/04/90)
In article <6264@wolfen.cc.uow.oz>, kirk@wolfen.cc.uow.oz (Kirk Barrett) writes: > G'Day, > has anyone had any experience with vmtune in the area of tuning a > Symmetry series to page instead of swap. We have am S27 with 8M of memory > and as soon as 2 or more people start running a certain statistics package > which uses a fair bit of memory, the system starts swapping, the load > average hits the roof (upto 21) and response time etc. drop considerably. > It has been suggested that I could tune the system to page instead of > swap with vmtune but that this may be tricky. > > Any Ideas or experiences ??? > Tuning the dynix vm system isn't that tricky. However, getting the result you want isn't always easy. To provide accurate advice, it would be usefull to know what the current vmtune parameters are and have some idea of your typical workloads. Are you using the monitor utility to examine the vm system? The vm system (with default settings) will prefer to page rather than swap. It will swap if the (page in+page out) rate > (40*the no. of swap devices) or if the amount of "free memory" drops below "des free". Check the monitor display. Without knowing a bit more about your environment, it's hard to make a suggestion, but consider these two scenarios: 1. A few users running assorted jobs plus your stats programs. The minRS value is low and large programs have little locality of reference. In this case you might expect high paging rates and hence begin swapping (if rate>40*spindles). In this case, it might be worth increasing minRS. You are likely to get fewer jobs in memory but more paging and less swapping. You should also monitor maxRS, to ensure that the larger jobs don't consume excessive amounts of real memory. 2. The stats jobs are essentially interpreters and have small text and large data sizes. maxRS is set relatively high and once again the locality of reference is poor. In this case, each process might quickly consume large amounts of real memory (upto maxRS) and free memory drops below desfree. Swapping begins. In this case, you could reduce maxRS to some point where the jobs are paging at a rate which provides acceptable response time. You would also need to juggle minRS so that sufficient copies could run and response time would also be a measure for determining this. During this exercise, you should remember that the vm system can/will adjust process's RSS between the limits minRS and maxRS depending on the paging rates. Refer to vmtune(8) for some extra information.