N-way, CMOS, and Master Processor Performance Considerations
Last Updated 3 April 1997
There are a few considerations to keep in mind when moving a VM/ESA system to a larger N-way configuration or a configuration where the individual processors are slower than the current system. This can be the scenario for certain migrations to new CMOS processors. For for details on VM/ESA and N-way configurations see the VM/ESA Greater N-way Thoughts presentation.
N-way Considerations
MP Factor - this refers to the efficiency losses in managing more and more processors. The VM/ESA MP Factor is very good, greater than 0.85 for all the scenarios I looked at and often greater than 0.90. This means that running VM/ESA on a 6-way would provide more than 85% of the capacity of 6 times the capacity of a single processor.
There is no UP gen for VM/ESA, unlike the 370 feature. This means there is a bit more overhead when running on a single processor, but provides for better efficiency in n-way configurations.
Areas limited by a Single Processor
In an n-way configuration, a non-MP work component is limited by the resources provided by a single processor. The following two cases illustrate this.
Case 1
Non-MP server virtual machine - the majority of server virtual
machines implemented in VM are not MP capable (that is if you
create a virtual MP machine with 2 or more virtual processors, the
application is not capable of running work on both processors at
the same time). This can be a limiting factor. For example, if
you have an SQL/DS server machine that makes up 25% of the workload
in terms of CPU cycles, it will be constrained on a 5-way processor.
On the 5-way, 20% (1 processor out of 5) is the max the SQL server
could access. In these cases, alternatives may include:
Case 2
Master Processor - one of the ways CP serializes work is with the
concept of a master processor. There is one and only one master
processor on a system (though it can change at times). Just like a
single non-MP server virtual machine, it is limited to a single
processor. So if 15% of the workload is master only, the system will
not fit comfortably on a 10-way where it is limited to 10%.
Measuring master processor is difficult. To bound the impact, collect
data on processor usage for master and alternates in the components
of Emulation time and non-emulation time (CP and System).
So master processor is between 4.6 and 6.2% of the workload, so the
system could run okay on a 15-way, but have problems on a 20-way.
In this type of conversion, make sure that the master and single
server virtual machine considerations listed above are checked.
Realize that instructions for individual jobs will take longer.
However, processor time is often just a small component of response
time.
Watch out for increases in other resource requirements because of
transactions running longer. For example, an application that is
somewhat compute bound will take longer to complete. Therefore,
the pages associated with that application will need to be in storage
longer. Moral: Provide sufficient resources to absorb any queueing
caused by slower processor.
Master non-emulation time
upper bound = ------------------------------
Total cpu time on all processors
Master non-emulation - avg(alternate non-emulation)
lower bound = ---------------------------------------------------
Total cpu time on all processors
For example, consider a 6-way processor with following CPU breakdown
Proc Total% Non-Emul% Emul%
0 95% 35% 60% ( master processor )
1 94% 8% 86%
2 94% 11% 83%
3 95% 8% 87%
4 93% 8% 85%
5 95% 9% 86%
upper = 35/(95+94+94+95+93+95) = 35/566 = 6.2%
lower = (35- ((8+11+8+8+9)/5))/566 = (35 - 8.8)/566 = 4.6%
Equal MIPS = more processors * slower processors
Back to the Performance Tips Page