Table of Contents
Advertisement
When a VM moves to a new node, VMware ESX immediately begins to migrate
its memory in this fashion. It adaptively manages the migration rate to avoid
overtaxing the system, particularly when the VM has very little remote memory
remaining or when the destination node has little free memory available. The
memory migration algorithm also ensures that it will not move memory
needlessly if a VM is moved to a new node for only a short period of time.
When all these techniques of initial placement, dynamic rebalancing, and
intelligent memory migration work in tandem, they ensure good memory
performance on NUMA systems, even in the presence of changing workloads.
When a major workload change occurs, for instance when new VMs are started,
the system takes time to readjust, migrating VMs and memory to new, optimal
locations. After a short period of time, the system completes its readjustments
and reaches a steady state.
Transparent page sharing optimized for NUMA
Many VMware ESX workloads present opportunities for sharing memory across
virtual machines. For example, several virtual machines might be running
instances of the same guest operating system, have the same applications or
components loaded, or contain common data. In such cases, VMware ESX
systems use a proprietary transparent page-sharing technique to securely
eliminate redundant copies of memory pages. With memory sharing, a workload
running in virtual machines often consumes less memory than it would when
running on physical machines. As a result, higher levels of overcommitment can
be supported efficiently.
Transparent page sharing for VMware ESX systems has also been optimized for
use on NUMA systems like the IBM x3950 M2. With VMware ESX running on a
multinode IBM x3950 M2 partition, pages are shared per node, so each NUMA
node has its own local copy of heavily shared pages. When virtual machines use
shared pages, they do not have to access remote memory.
Manual NUMA controls
Some administrators with advanced skills might prefer to control the memory
placement and processor use manually. See Figure 2-4 on page 71. This can be
helpful, for example, if a VM runs a memory-intensive workload, such as an
in-memory database or a scientific computing application with a large dataset.
Such an application can have performance improvements if 100% of its memory
is allocated locally, whereas VMs managed by the automatic NUMA
optimizations often have a small percentage (5-15%) of their memory located
remotely. An administrator might also want to optimize NUMA placements
manually if the system workload is known to be simple and unchanging. For
example, an eight-processor system running eight VMs with similar workloads
would be easy to optimize by hand.
70
Planning, Installing, and Managing the IBM System x3950 M2
Advertisement
Table of Contents
