Pages

Friday, September 14, 2018

IBM high-end POWER9-based servers set to dominate

By Bill Moran and Rich Ptak

 

IBM began shipping POWER9-based systems in 2017. IBM was determined to distinguish the offerings as unique by designing built-in capabilities specifically targeting emerging markets, i.e. Big Data, Deep Learning (DL) and AI. Today, some 80% of Fortune 100 companies have chosen IBM Power Systems to deliver the reliability and performance needed to support mission-critical applications and services.

On August 7, IBM introduced two new high-end POWER9 systems, the E950 and E980. We discuss the impact and implications of these servers, but first, some positioning and history.  

Why are Power Systems unique?

What is it that makes IBM Power Systems so unique? First, server vendors typically use commodity chips in their products. They have no direct control over the functions of the chip at the heart of their product. While they may exert some influence with the chip vendor[1], they use what is delivered, generic and unoptimized.

IBM owns the POWER9 chip and has end-to-end control over its overall system design. IBM can customize (optimize!) the microprocessor, the software[2], and all other hardware elements to target specific application areas. Designed-in capabilities provide a huge advantage for buyers developing or running AI, DL and Big Data applications.

IBM committed Power Systems to be an aggressively open, accessible platform. Thus, IBM'ssupport of the OpenPOWER initiative to develop an open eco-system dedicated to Power Systems. Participants can build their own POWER9-based systems and/or create enhancements that leverage and expand Power Systems strengths. The results are innovative extensions such as performance accelerators, i.e. OpenCAPI, NVIDIA NVLink, etc. In addition, there are long-term partnerships with major vendors, such as Google, SAP, TenCent[3], and PayPal. A recently announced performance-optimized SAP HANA platform  (discussed later) is just one significant solution resulting from such collaboration.

Another advantage results from IBM’s commitment to cloud-everywhere solutions. It’s a fact that today some 81% of enterprises have[4] a multi-cloud strategy. Monitoring and managing operations of multiple, diverse clouds can be complex and difficult. Therefore, IBM provides built-in cloud services to simplify the task. These include:  1) single console operations monitoring and management across multiple different clouds, 2) built-in PowerVM[5] to assure mobile security of virtualized workloads, and 3) IBM Cloud Management Console for Power Systems[6] that consolidates monitoring across multiple locations.

For 10 years running, IBM Power Systems have held the Number 1 spot as “most reliable system” with a 99.9996% uptime.

Unlike some vendors, IBM provides help for customers migrating across an evolving Power family in the form of cross-generation compatibility, extended services and support.  For example, IBM provides a temporary PowerVM[7] license to facilitate moving partitions from earlier systems to new POWER9 systems without requiring changes. Programs designed for POWER8 or POWER7 will run on POWER9. IBM’s Power Family roadmap includes POWER10. New and old Power Systems can combine in a single cluster. This assures older systems will perform well even as the datacenter environment changes.

Purchasing a computer invests you in its architecture and technology. Therefore, you benefit from knowing how customers making similar investments have fared during upgrades. IBM earns very positive endorsements in customer testimonials. Notably, IBM earned IBM X86 customer kudos for the support provided during the transition of IBM’s X86 servers to Lenovo.

Why care about all this? The roadmap illustrates IBM’s continuing commitment to POWER9. Extensive migration support demonstrates IBM protects customer investment by mitigating the cost and effort required to move to newer models.

A few statistics: for 10 consecutive years, IBM Power Systems have held the Number 1 spot as most reliable servers with 99.9996% uptime. An annual ITIC study[8] ranks the IBM mainframe, IBM Power, and Lenovo X86 as the three most reliable systems today. Lenovo acquired its X86 servers from IBM, therefore all three systems are IBM designs. One more key finding is that customers ranked IBM-provided service and support as outstanding.

IBM Power Systems – a complete family of servers

IBM announced additional POWER9 servers since its 2017 introduction. Our reviews of these earlier announcements are available here[9]. Table 1 summarizes today’s POWER9 family.
             Server Type                        Models                                    Focus Area     

Distributed
S914, S922, S924
Scale-out entry level to larger performance
Hyperconverged
CS821, CS822
POWER hardware + Nutanix software
Accelerated
AC922
POWER9 +NVidia GPU for max AI performance
Enterprise
E950, E980
Very Large Scale-up systems

Table 1 IBM's POWER9 Servers


IBM’s new E950 & E980 high-end servers

Today, the world’s most powerful computer, has an IBM POWER9 Microprocessor as its heart!

The newest extensions to the POWER9 family, E950 and E980, are highly scalable systems designed for large enterprise deployment. They preserve significant POWER family benefits such as support for on-prem, private, public, multi- and hybrid clouds. They also benefit from IBM’s demonstrated excellence at designing large systems as evidenced by the fact that the world’s most powerful computer has an IBM POWER9 microprocessor as its heart[10].

The new POWER9-based servers position the family to meet the needs of both scale-up and scale-out (distributed) infrastructure. Scale-up is needed for applications that achieve optimal performance by keeping the workload in a single memory space, like in-memory database apps. It delivers the practical efficiency of large system computing versus using many distributed systems.  Scale-out benefits applications involving a massive number of independent transactions[11], but quite low individual server utilization levels. Scale-up allows economically high utilization levels and the load balancing efficiency of simultaneously running many types of work.

Power Systems’ performance is well documented by ITIC (see above), customers, and IBM. IBM documents POWER9 advantages in key AI benchmarks[12]. At the chip level, POWER9 has demonstrated significant performance advantages over both earlier POWER chips and Intel’s X86 microprocessors.

In addition to sharing POWER9 characteristics, the E950 and E980 bring several unique advantages. First is the massive memory space of these systems. The E950 can attach up to 16 TB of memory; the E980 can attach a maximum of 64 TB. Eyes typically glaze over such numbers But, let’s dwell on them for a minute.

In-memory databases, like SAP HANA, have become quite popular. With 64 TB memory, it becomes possible to construct and contain in-memory a database of all living humans! Such massive databases will become increasingly common; the result of massive numbers of internet-attached devices. Big Data processing with reasonable performance now becomes practical for massive amounts of data.  

Benefits increase even more with the latest enhancements to IBM storage products. One more point, both new systems’ I/O subsystems can handle the millions of parallel I/O operations needed by AI solutions to process large amounts of disparate data.

Some business economics

Each enterprise is unique in its own ways. Therefore, any robust business case for a system requires customization beyond the scope of this paper. However, IBM provided a generic list price-based analysis of the cost of migrating from P7 to P9. It includes such common costs as software and hardware maintenance, list price comparison, etc. Overall savings exceed 50% when comparing an E950 (over 3 years) to a comparable POWER7-based system. IBM’s numbers and comparison appear to us to be accurate and valid. See Figure 1 below for the details. All amounts in $US.

IBM's analysis can help you to develop your own business case. Do note that there are additional enterprise-specific costs[13] to consider.



                                                                                                                                            Image courtesy of IBM, Inc.
Figure 1 3-year savings moving from P770 to E950

Our recommendation

It is a strong testament to IBM’s success at server design that some 80% of F100 companies chose IBM Power Systems to deliver the reliability and performance needed to support mission-critical applications and services.

IBM’s newest offerings both complement and extend the Power Systems family. We expect that experience with these latest, high-end, powerful systems will continue to attract and impress customers. We look forward to the results of head-to-head evaluations of POWER9 systems against its competitors.

Our analyses of POWER9, as well as the independent testimony to the system’s reliability and customer satisfaction ratings allows us to confidently recommend that anyone evaluating Linux systems give serious consideration to a POWER9-based system solution. IBM indicated that a POWER9-based follow-on to the AC922 is coming and plans for the POWER10 generation. We expect that both will be well worth commenting upon. We look forward to doing so.
  


[1] Typically, Intel or AMD.
[2] While IBM does not control Linux operating system design, they offer AIX (UNIX), and the IBM i OS on POWER9. Thus, they control these operating systems, and have demonstrated skill at adapting them to Power System hardware.
[3] A China-based global technology and investment company.
[4] See the 2018 RightScale State of the Cloud Report™ at https://www.rightscale.com/lp/state-of-the-cloud
[6] For information see: https://tinyurl.com/yajlyjkh
[7] PowerVM is standard on all POWER9 servers, no separate license required; full cloud support included.
[8] See https://tinyurl.com/y6upndcu. We recommend reviewing the entire report.
[11] The classic application model for this type is Google search. All transactions are independent of each other and there is no problem if the database is updated so that a search can show different answers at different times. Applications without these characteristics may have problems with distributed systems.
[12] https://www.ibm.com/blogs/research/2017/12/10x-faster-using-gpu/ includes insights into how IBM’s research team developed technology to speed up AI performance on POWER9.

[13] The IBM analysis excludes some environmental factors, e.g. new systems may require less floor space, or less power and cooling. The potentially financially significant impact is an enterprise-specific variable. Capital budget impacts (taxes, depreciation) were also excluded.