Deep Dive Into HP’s New HPC & Big Data Business Unit
When HP finally divides into two pieces – HP Inc. (PCs and printers) and Hewlett Packard Enterprise (servers and services) – how will the HPC portfolio fare? Views vary of course. The split is meant to let the ‘new’ companies shed distraction and sharpen focus. HPC will live within HP Enterprise, but perhaps surprisingly not by itself. Instead HPC is being combined with big data into a single global business unit, HPC & Big Data, created in March and led by longtime SGI executive and recent HP import, William Mannel.
“It’s not by mistake or coincidence we put HPC and big data together,” said Mannel, now VP and General Manager of the HPC & Big Data GBU at Hewlett-Packard. “We believe storing big data is one thing and we have technologies to do that. Getting productive use out of [the data] is another thing and many customers are using similar types of technologies to get value out of their big data.”
Hired last November and a key architect of the new GBU, Mannel discussed with HPCwire the HP strategy for expanding its HPC focus, why the timing is right to push into the enterprise, what some of the obstacles (and solutions) are, and the steady rise of new technologies – x86 still dominates (including at HP) but competitors (GPU, FPGA, OpenPower, ARM) are winning sockets in a trend likely to continue.
Don’t get the wrong idea. HP is hardly abandoning the HPC stratosphere. It still holds the overall lead in Top 500 systems (Nov. 2014) with 179 (36 percent) compared to IBM with 153 systems (30 percent) according to Top 500 organizers. It should be noted those numbers are down slightly for both companies. HP had 182 systems (36.4 percent) six months earlier, and IBM had 176 systems (35.2 percent).
“By any of the metrics you want to use we are already HPC leaders. We know we’re under represented in the top 100 and are expecting to drive forward into the portion of the market as well. However we think there’s additional opportunity to grow in the enterprise and that’s why we created a global business unit specifically to focus on HPC and big data.”
This notion of aligning HPC and big data has steadily gained traction. Many see a growing trend (or at least an obvious desire) by enterprises of all sizes to capitalize on big data (internally generated or externally available). Add to this the effort by vendors to evangelize and sell HPC to small and medium business seeking to differentiate themselves using traditionally HPC-dependent tools (modeling, simulation, etc.) and suddenly the enterprise HPC market looks enticing and big.
For the most part the analyst community seems to agree.
Market watcher IDC, which has adopted the HPDA acronym (high performance data analysis), has reported 67 percent of HPC sites use HPDA today and forecast the 2016 HPDA server and storage market at $1.2B and $800M respectively and growing faster than most segments.
Addison Snell, CEO, Intersect360, said, “HP’s strategy to combine HPC and big data internally is consistent with the industry dynamics we see, where there are now large categories of enterprise applications that are reliant on performance and scalability. One of HP’s strengths is its position in high-performance storage, and the company will want to leverage that. IBM has already started down this path, but the major enterprise storage vendors — NetApp, EMC, and HDS — are all missing it.”
“HP and Dell have both seen their share of HPC servers increase as a direct result of IBM, previously the clear #1, selling its x86 server business to Lenovo. While Lenovo will continue to sell into the HPC market, the long disruption to the IBM sales process opened up the door for competition. Now HP and Dell are nearly deadlocked for the HPC server market share lead,” said Snell, adding he thinks the HP split will further energize HP’s HPC efforts.
Apollo is the product line underpinning the HPC cum big data gambit. Introduced last June, the Apollo line spans supercomputing to the datacenter. The 8000 and 6000 were first to market, targeting the high end. Last month, HP announced the 4000 (big data) and 2000 (entry level, datacenter) additions to the line.
Outside the enterprise, HP has already racked up impressive wins. The first implementation of the Apollo platform (8000 machines) was by DOE’s National Renewable Energy Laboratory (NREL). NREL worked with Intel and HP to build Peregrine, a warm-water, liquid-cooled supercomputer. (The warm water is reused to heat the building after cooling the computer.)
Peregrine has 6,912 Intel Xeon E5-2670 “SandyBridge” processor cores, 24,192 Intel Xeon E5-2695v2 “IvyBridge” processor cores for a total of 31,104 Intel Xeon processor cores, providing a total of about 608 TeraFLOPS or Trillion floating point calculations per second. Peregrine also has 576 Intel Phi many-core co-processors with an aggregate performance of about 582 TeraFLOPS. In total Peregrine is capable of 1.19 PetaFLOPS.
In April of this year, AGH University of Science and Technology in Krakow brought online its Prometheus supercomputer. Packing 1.7 Petaflops of peak computational performance, the HP-built machine is the most powerful supercomputer in the history of Poland and the world’s largest installation of HP Apollo 8000 servers. The 30 metric ton machine houses 1,728 HP Apollo 8000 InfiniBand-connected servers inside 15 racks.
High profile wins such as these, HP hopes, will create a buzz around the entire Apollo line including its more recent members targeted at big data and datacenter activities. Here is a brief snapshot of the Apollo platform:
- Apollo 2000 is the entry offering. It’s available with up to four servers in 2U chassis, uses Intel Xeon ES-2600 processors, and supports as many as 24 drives per node. “You can use one as the head node and the other three as computes nodes,” says Mannel.
- Apollo 4000 (three systems in the line, 4200, 4530, 4510) is aimed squarely at Big Data and the datacenter. Mannel noted, “It’s a big data Platform specifically used for matching compute with a lot of storage. It’s not a RAID box but it is a storage server with a number of different configurations.”
- Apollo 6000 & 8000 as noted earlier target large-scale systems in technical and scientific computing. The users, according to Mannel, need hundreds to thousands of cores. Xeon E3 and E5 processors are used throughout the line and top models have two accelerator slots that support Xeon Phi.
Time will tell if this is the right product mix. Currently much of the enterprise market is sluggish. HP’s most recent financial results, released May 22, revealed Enterprise Group revenue was down 1% year over year with a 14.5% operating margin. Industry standard servers revenue was up 11%, but storage (8%), business critical systems (15%), networking (16%) and technology services (8%) were all down. Likewise, Enterprise services (16%), infrastructure technology outsourcing (20%), and application and business services revenue (8%) all declined.
Market fluctuations aside, the low hanging fruit would seem to be dual-use opportunities in large industries where HPC is already established such as the auto industry.
“Buy an auto today, the thing itself is a big data producer. It uploads all this data, which gets collated and collected and analyzed from quality standpoint from a driver preference standpoint. Many of the big auto manufacturers have big data projects at the same time they are using their HPC resources for more of the standard structural analysis and crash analysis and fluid dynamics types of analysis,” said Manuel.
Making the HPC & big data gambit work for companies less experienced in HPC and with fewer computational resources will be challenging. For starters, adopting HPC isn’t easy. Complicated systems management, new programming techniques, tricky application software, power & space requirements, and unfamiliar architectures can quickly confound new-to-HPC users.
HP understands the challenges, contended Mannel, and has an effective strategy.
“One core bottleneck is that HPC has been so generic. What’s needed is a solution approach in which the right application with right configuration of the hardware and the right level of management and usability that make it acceptable to customers. I think that’s one area where the HPC market has struggled,” he said.
Within the HPC and big data GBU is a formal HPC Pursuit Group, which includes a team of applications engineers who work with end user customers, ISVs, and the open source community to ensure applications run well on HP hardware and for the customer’s specific application and workload. The group works with customers pre- and post sale to “make sure they are getting the performance they want.”
It’s also important to offer alternative access routes to HPC resources (e.g. cloud) said Mannel both for pilot programs and production environments.
“We provide for a number of very large customers as well as smaller customers the ability to essentially get HPC on tap. If they don’t want to build their own HPC datacenter because of space constraints or lack of expertise, we’ll set up customers with access to HPC resources we control. It can be done on or off customer premise; they just pay a monthly bill,” Mannel said.
For now, HP’s key target markets are pretty vanilla:
- Oil & Gas, a long-term HPC user and frequent adopter leading advances;
- Manufacturing, “[HP] has expertise in computing and engineering work which tends to be engineering simulation more than anything else;”
- Financial services, again no surprise;
- Life sciences, which “is an emerging market for us. It’s a place where we are going to put more investment and expertise over time. It was a past focus but we are recommitting to it.”
Currently x86 technology dominates the HPC portfolio but its preeminence is being slowly chipped away. In terms of balancing the portfolio by incorporating various technologies, there is a shift going on – not that x86 is less important but that other technologies are also becoming important and taking a greater share.
“That’s definitely fair,” said Mannel. “You’re seeing a massive variety of different technologies and techniques coming onboard. Here’s an example. Nearly ten years ago I led a big effort around using FPGAs for HPC. We had a little success but in the end it was only for a few defense applications. It was just hard to program FPGAs.
“Come forward ten years and I’ve talked to some service providers who are swearing by using FPGAs. They are saying, ‘We did the [programming] work, are getting acceleration, and are very happy with the results.’
“I think that’s a really common experience. Now there’s GPUs out there, Intel Phi, high-performance interconnects, new programming approaches. It’s all coming about because just waiting for the latest rev of the [next Intel] chip is not creating the level of performance, price performance, and performance per watt that customers are expecting,” noted Mannel
Intel haters should not get giddy. No sea change is expected soon, certainly not at HP, but change is now part of the technology selection conversation. An HP example is its Moonshot system aimed at cloud and more traditional datacenter functionality. Moonshot has a power-stingy system-on-chip ARM processor in its arsenal. Along with other technology, the ARM processor allows Moonshot to have a small space and small power footprint.
“I use a phrase, ‘The right compute at the right time for right data.’ For a long time, the x86 was just that. You could get a very wide range of applications that ran very well on x86 architecture. Now you are starting to see with some of these new applications that you can get better performance using different technologies. We tend to invest in new technologies that fit our capabilities as a company, our drivers, and also offer strong value proposition for customers.”