As customers begin to deploy z/VM 6.4, they might wish to give consideration to the following items.
Our regression findings included results from both memory-rich and memory-constrained workloads. The memory-constrained workloads tended to get improvements in ETR and ITR, especially if they were large N-core configurations. This was because of the compounding of the effects of the paging improvements and the memory scalability improvements. If your workloads are memory-constrained and large N-core, IBM would appreciate hearing your experience; please send us a feedback.
z/VM 6.4 lets the system administrator switch the system between SMT-1 mode and SMT-2 mode without an IPL. In this way the administrator can try SMT-2, measure its behavior, and then return to SMT-1 mode if SMT-2 mode is found unsuitable.
Customers must remain mindful that SMT-1 mode is not the same as non-SMT mode. IBM did measure the performance of z/VM 6.4 in SMT-1 mode compared to non-SMT mode, core counts being equal. Differences were slight if visible at all. However, the total computing capacity achievable in non-SMT mode still exceeds, and will forever exceed, what can be achieved in SMT-1 mode or SMT-2 mode. In non-SMT mode z/VM can make use of 64 cores in the LPAR whereas in SMT-1 mode or SMT-2 mode z/VM can make use of only 32 cores. Customers running z/VM in non-SMT mode in LPARs having more than 32 cores will have to give up some cores to get to SMT-1 or to SMT-2. Switching z/VM between non-SMT mode and one of the SMT modes still requires an IPL.
In using Dynamic SMT to audition SMT-2, please remember to collect reliable measurement data and to use your data to drive your decision about how to proceed. While you are in SMT-1 mode, collect both application-specific performance data, such as transaction rates, and MONWRITE data. Be sure to collect CPU MF data as part of your MONWRITE collection. Then switch to SMT-2 and collect the very same data. Then compare the two sets of results. Try to compare results collected during times when workload was about the same. For example, for your situation it might make sense to compare Tuesday at 2 PM this week to Tuesday at 2 PM last week.
IBM is interested to hear the experiences of customers who use Dynamic SMT to audition SMT-2. Please take time to send us a feedback. We will be grateful if you will let us have copies of your measurement data and of your comparison analysis.
HyperPAV and zHPF Paging
In our article we discuss the z/VM 6.4 changes that let CP use HyperPAV aliases for paging I/O. We also discuss the changes that let CP use High Performance FICON (zHPF) for paging I/O.
Achieving paging I/O concurrency is one reason customers have been asking for paging I/O to exploit HyperPAV aliases. Customers who have defined large numbers of paging volumes only for the purpose of achieving I/O concurrency can now use HyperPAV aliases to achieve the concurrency. Doing this will let them return the excess paging DASD to other uses. In doing this remember not to decrease the total amount of paging space below a safe level for your workload.
Performance Toolkit for VM has not yet been enhanced to depict what the monitor records report about CP's use of HyperPAV aliases. In the meantime there is a VM Download Library package called HPALIAS you can use.
The work done in the memory scalability improvement let IBM increase the central storage support limit from 1 TB to 2 TB. The increase will help customer who are feeling memory constraints. The heart of the work was to split the memory manager state data from one monolithic structure into a structure that could be used concurrently by multiple logical CPUs without undue spin lock contention.
Customers wanting to increase memory need to remember that very often a system can grow effectively only when all resource types are increased in proportion to one another. CPUs, memory, channels, networking, paging DASD space, and paging DASD exposures are some examples of resources that need to be considered together when planning system growth.
CP Scheduler Improvements
In our article we discuss the improvements made in the CP scheduler in z/VM 6.4. The purpose of the improvements was to address VM65288, in which a customer demonstrated CP did not honor share settings in certain situations. In the tests we tried, CP now honors share settings more accurately than it did in previous z/VM releases.
Some customers might have finely tuned their systems by adjusting share settings until the system behaved just the way they wanted. Customers who have done so might find they need to retune now that the scheduler has been repaired.
On systems that are not completely busy, in a very large, macro sense, share settings in theory do not matter. As a consequence, unused capacity in the LPAR is often taken as a sign that all users are getting all of the CPU time they want. During our study of z/VM 6.4, though, we found that when there are only a few users and their demands add up to a good fraction of the LPAR's capacity, there can still be unfulfilled demand even though the LPAR is not completely busy. The Perfkit reports FCX114 USTAT, FCX114 INTERIM USTAT, FCX164 USTATLOG, and FCX315 USTMPLOG can help you to find users who want more CPU power than they are being given. FCX135 USTLOG is less useful because it does not cite users individually.
In the scenarios we tried, z/VM 6.4 often dispatched users with less delay than did the previous release. In other words, the amount of time between the instant a virtual CPU became ready for dispatch and the instant CP dispatched the virtual CPU tended to be less on z/VM 6.4 than it was on the previous release. This bodes well for workloads where response time is important or for situations where a physical phenomenon, such as a network device interrupt, must be serviced quickly to avoid a problem such as a timeout or a data overrun.
During our study we also evaluated the LIMITSOFT and RELATIVE LIMITHARD features of the CP scheduler. We found those two features not to work correctly in the scenarios we tried. Customers depending upon LIMITSOFT or RELATIVE LIMITHARD might wish to evaluate whether their use of those features is having the intended effect.