As customers begin to deploy z/VM 6.2, they might wish to give consideration to the following items.
On Load Balancing and Capacity Planning
z/VM Single System Image offers the customer an opportunity to deploy work across multiple partitions as if the partitions were one single z/VM image. Guests can be logged onto the partitions where they fit and can be moved among partitions when needed. Movement of guests from one member to another is accomplished with the VMRELOCATE command.
Easy guest movement implies easy load balancing. If a guest experiences a growth spurt, we might accommodate the spurt by moving the guest to a more lightly loaded member. Prior to z/VM 6.2, this kind of rebalancing was more difficult.
To assure live guest relocation will succeed when the time comes, it will be necessary to provide capacity on the respective members in such a manner that they can back each other up. For example, running each of the four members at 90% CPU-busy and then expecting to be able to distribute a down member's entire workload into the other three members' available CPU power just will not work. In other words, where before we tracked a system's unused capacity mainly to project its own upgrade schedule, we must now track and plan members' unused capacity in terms of its ability to help absorb work from a down member. Keep in mind that members' unused capacity is comprised of more than just unused CPU cycles. Memory and paging space also need enough spare room to handle migrated work.
As you do capacity planning in an SSI, consider tracking the members' utilization and growth in "essential" and "nonessential" buckets. By "essential" we mean the member workload that must be migrated to other members when the present member must be taken down, such as for service. The unused capacity on the other members must be large enough to contain the down member's essential work. The down member's nonessential work can just wait until the down member resumes operating.
When one partition of a cluster is down for service, the underlying physical assets it ordinarily consumes aren't necessarily unavailable. When two partitions of an SSI reside on the same CEC, the physical CPU power ordinarily used by one member can be diverted to the other when the former is down. Consider for example the case of two 24-way partitions, each normally running with 12 logical PUs varied off. When we take one partition down for service, we can first move vital guests to the other partition and simultaneously vary on logical PUs there to handle the load. In this way we keep the workload running uninterrupted and at constant capacity, even though part of our configuration took a planned outage.
Sometimes achieving work movement doesn't necessarily mean moving a running guest from one system to another. High-availability clustering solutions for Linux, such as SuSE Linux Enterprise High Availability Extensions for System z, make it possible for one guest to soak up its partner's work when the partner fails. The surviving guest can handle the total load, though, only if the partition on which the guest is running has the spare capacity, and further, only if the surviving guest is configured to tap into it. If you are using HA solutions to move work, think about the importance of proper virtual configuration in achieving your goals when something's not quite right.
On Guest Mobility
The notion that a guest can begin its life on one system and then move to another without a LOGOFF/LOGON sequence can be a real game-changer for certain habits and procedures. IBM encourages customers to think through items and situations like those listed below, to assess for impact and to make corrections or changes where needed.
Charge-back: Can your procedures for charge-back and resource billing account for the notion that a guest suddenly disappeared from one system and reappeared somewhere else?
Second-level schedulers: Some customers have procedures that attempt to schedule groups of virtual machines together, such as by adjusting share settings of guests whose names appear in a list. What happens to your procedures if the guests in that group move separately among the members in an SSI?
VM Resource Manager: VMRM is not generally equipped to handle the notion that guests can move among systems. IBM recommends that moveable guests not be included in VMRM-managed groups.
On MONWRITE and Performance Toolkit
There continues to be a CP Monitor data stream for each of the individual members of an SSI. To collect a complete view of the operation of the SSI, it will therefore be necessary for you to run MONWRITE on all of the members of the SSI. Remember to practice good archiving and organizing habits for the MONWRITE files you produce. During a performance tuning exercise you will probably want to look at all of the MONWRITE files for a given time period. If you contact IBM for help with a problem, IBM might ask for MONWRITE files from all systems for the same time period.
Performance Toolkit for VM continues to run separately on each member of the cluster. There will be a PERFSVM virtual machine on each member, achieved through the multiconfiguration virtual machine support in the CP directory.
Now more than ever you might wish to configure Performance Toolkit for VM so that you can use its remote performance monitoring facility. In this setup, one PERFSVM acts as the concentrator for performance data collected by PERFSVM instances running on other z/VM systems. The contributors forward their data through APPC/VM or other means. Through one browser session or one CMS session with that one "master" PERFSVM, you as the performance analyst can inspect data pertaining to all of the contributing systems.
Performance Toolkit for VM does not produce "cluster-view" reports for resources shared among the members of an SSI. For example, when a real DASD is shared among the members, no one member's MONWRITE data records the device's total-busy view. Each system's data might portray the volume as lightly used when in aggregate the volume is heavily busy. Manual inspection of the individual systems' respective reports is one way to detect such phenomena. For the specific case of DASD busy, the controller-sourced FCX176 and FCX177 reports might offer some insight.
On Getting Help from IBM
If you open a problem with IBM, IBM might need you to send concurrently taken dumps. Be prepared for this. Practice with SNAPDUMP and with PSW restart dumps. Know the effect of a SNAPDUMP on your workload. Be prepared for the idea that you might have to issue the SNAPDUMP command simultaneously on multiple systems. Practice compressing dumps and preparing them for transmission to IBM.
On the CPU Measurement Facility Host Counters
Starting with VM64961 for z/VM 5.4 and z/VM 6.1, z/VM can now collect and log out the System z CPU Measurement Facility host counters. These counters record the performance experience of the System z CEC on such metrics as instructions run, clock cycles used, and cache misses experienced. Analyzing the counters provides a view of the performance of the System z CPU and of the success of the memory cache in keeping the CPU from having to wait for memory fetches. The counters record other CPU-specific phenomena also.
To use the new z/VM CPU MF support, do the following:
- Run on System z hardware that includes the
CPU Measurement Facility. A z10 at driver 76D bundle 20 or
later, or any z196, or any z114 is all that is needed.
- Run z/VM 5.4 or z/VM 6.1 with VM64961 applied,
or run z/VM 6.2.
- Authorize the z/VM partition to collect its
partition's counters. This is done at the System z
SE or HMC. The IBM red paper
Setting Up and Using the IBM System z CPU Measurement Facility with
describes this step. See section 2.2.2 therein for details.
The steps to authorize a partition at the SE or HMC
are the same regardless of whether the partition runs z/VM or
Configure CP Monitor to emit sample records in
the PROCESSOR domain. Most customers running Monitor already
configure CP Monitor in this way.
To check whether Monitor is already doing this, just issue
CP QUERY MONITOR
and look at the output to see whether
processor samples are enabled.
If they are not, place a
CP MONITOR SAMPLE ENABLE PROCESSOR command at the
spot in your startup automation that turns on Monitor.
This might be MONWRITE's PROFILE EXEC or PERFSVM's
PROFILE EXEC, for example.
Start recording MONWRITE data as you normally would.
- To check that everything's working, issue CP QUERY MONITOR. Look for processor samples to be enabled with the CPUMFC annotation appearing in the output. If you see NOCPUMFC, something's gone wrong. Check everything over, and if you can't figure it out, contact IBM for help.
Once these steps are accomplished, the new CPU MF sample records, D5 R13 MRPRCMFC, will appear in the Monitor data stream. MONWRITE will journal the new records to disk along with the rest of the Monitor records. Performance Toolkit for VM will not analyze the new records, but it won't be harmed by them either.
While it is not absolutely essential, it is very helpful for MONWRITE data containing D5 R13 MRPRCMFC records also to contain D5 R14 MRPRCTOP system topology event records. Each time PR/SM changes the placement of the z/VM partition's logical CPUs onto the CPC's CPU chips and nodes, z/VM detects the change and cuts a D5 R14 record reporting the new placement. For the D5 R14 records to appear in the monitor data stream, the system programmer must run CP Monitor with processor events enabled. Note also that the D1 R26 MRMTRTOP system topology config record is sent to each *MONITOR listener when the listener begins listening. APAR VM64947 implements the D5 R14 records on z/VM 5.4 or z/VM 6.1.
IBM wants z/VM customers to contribute MONWRITE data containing CPU MF counters. These contributed MONWRITE files will help IBM to understand the stressors z/VM workloads tend to place on System z processors. For more information about how to contribute, use the "Contact z/VM" link on this web page.
Prior to z/VM 6.2, IBM offered a z/Architecture-mode CMS, called z/CMS, as an unsupported sample. In z/VM 6.2, z/CMS is now supported. Some customers might consider z/CMS as an alternative to the standard ESA/XC-mode CMS, which is also still supported.
z/CMS can run in a z/Architecture guest. This is useful mostly so that you can use z/Architecture instructions in a CMS application you or your vendor writes. A second point to note, though, is that a z/Architecture guest can be larger than 2 GB. z/CMS CMSSTOR provides basic storage management for the storage above the 2 GB bar, but other CMS APIs cannot handle buffers located there.
If you use z/CMS, remember that z/Architecture is not ESA/XC architecture. A z/CMS guest cannot use ESA/XC architecture features, such as VM Data Spaces. This means it cannot use SFS DIRCONTROL-in-Data-Space, even though the SFS server is still running in ESA/XC mode. Similarly, one would not want to run DB/2 for VM under z/CMS if one depended on DB/2's MAPMDISK support.
If you are using an RSK-exploitive application under z/CMS, remember that the RSK's data-space-exploitive features will be unavailable.