> I've always been under the impression that the best configuration for > Linux/390 guest swap is to make the virtual machine storage > as large as it > SHOULD need under normal operating conditions, and give the guest > practically NO swap.
Not quite. The idea is to keep the virtual machine size as small as possible, *and* allow Linux to swap via the most efficient method possible if it's necessary. Linux dies horribly if it hits a burst of activity and can't find enough swap to handle the load. > Since Linux wants to "exploit" his > memory and swap > space as if he were a "stand alone" machine - the intent is > to not give him > more than he absolutely needs to prevent him from doing > "clever things" with > it, no? True for virtual memory. Swap gets used when no virtual memory is available, so you need to make sure that if you constrain the machine by reducing virtual machine size, then you also make sure that the virtual machine has sufficient swap to survive a sudden demand for virtual storage within the Linux machine. Since you need usage information to size that, if you start with a relatively large amount of swap and a small virtual machine (let's say, using my rule of thumb, a 64M virtual machine, and 142M of swap), then you should be able to quickly determine what the average swap utilization will be under load. You can then either adjust the virtual machine size to avoid swapping, or you can accept some amount of swapping and keep the maximum working set size for the virtual machine small, and expend the real storage on VDISK for swap, which gets allocated only if it's really needed. I'm sure that others will have different views, but I tend to try for solutions that allow CP to do as much optimization as possible without Linux having to know or care about it, and I think that allowing the Linux guests to swap to VDISK is overall somewhat more efficient use of resources, as VDISK pages are pageable, and CP can coordinate that process much more efficiently than Linux does. Increasing the virtual machine size tends to make CP scheduling more awkward in that it essentially permanently increases the WSS committed to that virtual machine (and allows Linux to try to be clever, but end up failing and increasing the WSS to the maximum possible for that virtual machine size). VDISK is allocated only when actually needed, released when not in use, and is a pooled resource that CP can manage across virtual machine boundaries. Another note: note that the 2x+32M is a *starting* figure, which gets adjusted over time. If you start with too much, you'll be able to get a better read on what you really need w/o risking a crash-and-burn scenario if you run out. This will never be a "set and forget" deal -- you'll need frequent measurement to keep this balanced. In a lot of real cases, if I have a 64M machine, I rarely end up with more than about 16M of swap based on actual measurement of what the applications *do*, but I have no way of knowing that until I get some real data on it. On the other hand, if I know the machine will be running a piggy app like WAS which will have a large WSS, then the extra virtual machine size is probably worth it (and that machine will probably need the swap anyway...8-)). -- db ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
