|< Day Day Up >|| |
Page space has two requirements: sufficient space, and enough devices. The space calculation is very simple: add up all of the virtual machine sizes you will support, add up all of the virtual disks that will be used, and multiply by 4. This is conservative, but will ensure that you do not run out of storage.
The multiplier of 4 is for performance reasons. As available page space decreases, so will the block size, and then the I/O requirement will go up accordingly. In practice, if some amount of attention is paid, a multiplier of 2 will provide the same performance. But then if several Linux servers are added, you might be down to 1 or less before an opportunity to analyze the page space is taken.
The device calculation is based on workload and device technology. A target maximum device busy of 20% is reasonable. If you are paging to the RVA devices on ESCON channels (like those in Figure 12-6 on page 304) with 15 to 20ms response time, then each device can support about 10 I/Os per second.
However, if the page devices are on ESS with FICON channels with 1 to 2ms response time, then each device will support over 100 I/Os per second. Since we had no opportunity to run tests in a storage-constrained environment, there was very little paging.
There is a trade-off between having a Linux swap disk on real versus virtual disk. When the swap space is on a real disk, it will be a single point of contention. The sustained swap rate can reach 100 or more, depending on your DASD. At this point, the Linux server will have very poor performance. When the Linux swap space is on virtual disk, there is no additional resource cost until the virtual disk is used.
If there were 50 servers, each with 2 GB of real disk allocated for swap, there is a 100 GB requirement. Since each of those disks are dedicated to a server, all swap activity from one server will go to one disk.
The other option is to define those 50 disks as page devices. Under times of very heavy swapping, and when z/VM does need to offload pages, the I/O is now spread over the 50 page devices. This increases the total real I/O bandwidth potential by each individual server to 50 times what could be supported on just one device. The total supportable load for the system has increased significantly, and provides the ability for each individual Linux guest to page at very high rates with no ill effect on performance. This is a tremendous advantage of virtual disk over any other swapping technology.
|< Day Day Up >|| |