ScaleMP The Dawn of the Virtualized Data Center

milkygoodyearAI and Robotics

Oct 14, 2013 (4 years and 9 months ago)


© The Bloor Group


The Dawn of the Virtualized Data Center
by Robin Bloor,
Chief Analyst
The Common Sense of Virtualization
When it became clear in about 2006 that VMware had established a market for
software that could partition a single server into a handful virtual servers, it occurred
to us that it would not be long before VMware itself, or some other company, moved
up in scale and began to offer technology that virtualized clusters or even grids of
servers. It has happened, although not as swiftly as we expected it to, and the
company that is leading the charge is ScaleMP.
The virtue in creating virtualized environments is clear. The rapid escalation in x86
computer power had made it possible to run multiple applications on single
commodity servers. Unfortunately, on the x86 platform we did not have operating
systems (OSs) that could easily manage multiple servers at once. The x86 OSs that
became dominant, Windows and Linux, were not built for that. But it didn’t matter
too much because VMware appeared on the scene just as the excessive power of
commodity servers was beginning to become an embarrassment. It enabled data
centers to parcel up such power into multiple virtual machines running on a single
server. Windows and Linux then added such virtualization capabilities.
So partitioning the individual resources of x86 servers thus became fairly easy, but
aggregating such servers into clusters or grids was not so easy. This is where
ScaleMP stepped in.
Scale-Up Virtualization and Big Data
Put simply, ScaleMP aggregates many x86 systems to become a single virtual x86
symmetric multiprocessor (SMP) computer. The virtue of being able to do this is not
obvious to you, it becomes so as soon as you think of the words “Big Data.”
Incidentally, while we may tend to think of Big Data as being characterized by the
need to process very large volumes of data, this is over-simplistic. “Big Data”
workloads can be very data intensive but they can also be very compute intensive.
This can certainly be the case, for example, when employing machine learning
As regards aggregation of resources, one of the virtues of ScaleMP is that it can
configure very large virtual machines that are suited, in respect of the resources they
provide, to the workload. It is also possible to run virtual machines within a ScaleMP
virtual machine. So for example, you could configure a six node cluster and allocate
the resources of 1.5 nodes to each virtual machine within the virtual cluster.
You could, if you so wished, use ScaleMP to aggregate a collection of servers and use
them as a private cloud for a collection of applications running 24x7. At the same
time another collection of servers could be configured for much more resource-
intensive data analytics applications. As the resource demands changed for either
collection of servers, resources could be swapped from one to the other. However, in
practice, ScaleMP is used in a more sophisticated manner, optimizing the whole x86
resource space, according to varying workloads.
The effect of deploying ScaleMP within a data center is illustrated above. The first
point to note is that there is no need to deploy specialist hardware to cater for known
large workloads. That fat node cluster and the shared memory system illustrated in
the Legacy Data Center (left) is virtualized in the On-Demand Data Center (right).
Commodity servers become basic building blocks and incremental increases in
resource needs do not lead to expensive addition hardware purchases.
It is worth emphasizing here that ScaleMP is a software-only solution. Its Versatile
SMP (vSMP) architecture has been designed to deliver the functionality of proprietary
chipsets via software and to utilize commodity interconnects such as InfiniBand. The
virtual machine itself only uses a small fraction of any node’s processor and memory
resources. It runs on industry-standard off-the-shelf components, providing the cache
coherency, shared I/O and system interfaces (BIOS, ACPI) that are needed by any
underlying OS. There are no additional device drivers needed and no alterations are
required to either the OS any application running within any OS that it hosts.
ScaleMP Options
ScaleMP is available in two variants: vSMP Foundation and vSMP Foundation
Advanced Platform. vSMP Foundation provides the basic ScaleMP capability,
enabling aggregation of up to 32 servers into one virtual SMP system. Prices for this
currently start at $400 per socket. This is aimed at customers that wish to configure
one or more large SMP clusters on inexpensive commodity hardware.
The vSMP Foundation Advanced Platform is intended for data center wide use and
delivers the full capability. It provides an on-demand SMP capability that can
aggregate up to 128 servers into a single virtual machine. Such VMs can be
provisioned by traditional provisioning systems (Bright Cluster Manager, Insight
CMU, etc.) The partitioning capability within the Advanced Platform enables the
© The Bloor Group


The Dawn of the Virtualized Data Center
“It is worth
here that
ScaleMP is a
software only
...It runs on
standard off-
ScaleMP’s Virtualized Data Center
creation of hierarchies of virtual machines within any given SMP cluster in a flexible
manner and it includes tools for resizing system partitions according to changing
resource demands. This also delivers Parallel InfiniBand performance, supporting up
to 4 parallel InfiniBand links, to provide up to 224GBps per node using InfiniBand
FDR (56GBps). Prices for this current start at $800 per socket.
In total the vSMP Architecture can currently support up to 256TB of memory and up
to 32,768 processor cores.
The Business Advantages
Ultimately what ScaleMP delivers is the ability to turn a farm of hundreds of x86
servers into a managed resource space. While the circumstances of each large data
center varies, the ability to harmonize the x86 resources and manage them as a single
unit is, on its own, compelling and the use of ScaleMP’s vSMP Foundation will likely
lead to reduced labor costs in many areas and reduced complexity in managing the
One can think of ScaleMP as delivering an out-of-the-box server consolidation
capability, so its initial implementation could have a significant impact on operating
expenses (OPEX) in many ways from reducing power and cooling requirements to
reducing floor space.
By virtue of removing the need for specialized hardware it will reduce capital
expenditure for those sites that run resource-hungry applications that demand either
considerable memory resources or considerable computer resources. In general, given
that it allows data centers to standardize on commodity hardware components, it is
likely to make hardware procurement much simpler as well as less expensive.
But ultimately, the overall business benefit is in the flexibility that it can deliver. A
company’s computing needs change constantly and sometimes dramatically. Because
ScaleMP delivers deployment and configuration flexibility, it will be able to
accommodate major changes in resource requirements.
A Simpler Way
Recently, IBM has teamed with ScaleMP to offer two systems that transparently scale
to new heights, offering tremendous amounts of resources in a single system. Two
separate appliances have been developed that easily integrate into a data center

IBM Very Large Memory Appliance

An optimized appliance with the ability
to address up to 7.5TB of memory. This large-memory appliance series is
capable of handling very large workloads that require significant amounts of
data. The solution is based on IBM System x3750 M4 and helps speed time to
solution for applications in areas such as analytics, in-memory databases and

Virtual Cluster Appliance

Eliminate day-to-day complex cluster management
tasks for midrange clusters with up to 224 cores and 1.75TB RAM. The
solution is based on IBM BladeCenter H with IBM HS23 blades and provides
© The Bloor Group

“In our view,
ScaleMP is a
whose time has

The Dawn of the Virtualized Data Center
an extremely powerful computing platform driven by a single OS instance,
hence managed as a single server.
Preconfigured and pretested large memory configurations from a leading supplier
like IBM are a definite way to speed the deployment into a flexible data center. These
systems simplify the overall process and can lead to a greater ROI.
The Bottom Line
It does not surprise us at all that ScaleMP has been gaining considerable traction
among companies with large data centers. It delivers a level of manageability to x86
users that has previously and sadly been absent. In our view, ScaleMP is a technology
whose time has come.
© The Bloor Group

Company: ScaleMP
vSMP Foundation and
vSMP Foundation Advanced Platform