Future Computing Platforms for Science in a Power Constrained Era
Power consumption will be a key constraint on the future growth of Distributed High Throughput Computing (DHTC) as used by High Energy Physics (HEP). This makes performance-per-watt a crucial metric for selecting cost-efficient computing solutions. For this paper, we have done a wide survey of current and emerging architectures becoming available on the market including x86-64 variants, ARMv7 32-bit, ARMv8 64-bit, Many-Core and GPU solutions, as well as newer System-on-Chip (SoC) solutions. We compare performance and energy efficiency using an evolving set of standardized HEP-related benchmarks and power measurement techniques we have been developing. We evaluate the potential for use of such computing solutions in the context of DHTC systems, such as the Worldwide LHC Computing Grid (WLCG).
1 Introduction and Motivation
The data produced by the four experiments at the Large Hadron Collider (LHC)  or similar High Energy Physics (HEP) experiments requires a significant amount of human and computing resources which cannot be provided by research institute or even country. For this reasons the various parties involved created the Worldwide LHC Computing Grid (WLCG) in order to tackle the data processing challenges posed by such a large amount of data. The WLGC consists of a highly federated union of computing centers sparse in 40 countries and it represents an admirable example of international organization. The Compact Muon Solenoid (CMS) experiment  at the LHC alone uses the order of 100,000 x86_64 cores for its data analysis and similar happens for the other general purpose LHC experiment, ATLAS. Moreover, as the LHC and the experiments will undergo planned luminosity upgrades over the next 15 years , its dataset size will increase of 2-3 order of magnitude, which will require additional efforts to increase its processing capacity.
In this paper we continue our ongoing effort to explore the performance and viability of various computing platforms which we consider some of the most likely components of tomorrow data centers. We asses their performance based on both synthetic benchmarks and realistic workflows used in production at CMS.
2 Considered platforms
For this study we have selected five major product lines:
Intel Xeon: Intel is the leading CPU manufacturer in the world. Its Xeon product line is its flagship brand and provides solutions for all ranges of computing centers. While the Xeon brand is usually associated to highly performant, architecturally advanced CPUs, over the years Intel made sure to take into account power efficiency needs. This is reflected by the inclusion of features, like SpeedStep and TurboBoost, specifically aimed at a more power efficient usage or RAPL which is used to monitor CPU power usage itself.
APM XGene: ARM architecture based products are, when considered as a single entity, market leader for low power CPUs, in particular due to their wide spread adoption in a wide range of consumer electronics product like cellular phones and tablets. Thanks to the economy of scale of cell phones aims to become a serious player in the server market, in particular Applied Micro (APM) X-Gene product line is one of the first attempts at providing a 64-bit ARMv8 chip which is suitable for the low power, high density server market. We have already detailed the main difficulties sustained to port to such an architecture in a preceding work .
Intel Atom: Intel Atom architecture is Intel solution for the power efficient market. It consists of a standard x86_64 core, where particular trade-offs have been made to reduce complexity, sacrificing performance for power efficient. E.g. Atom processors has a simpler, in order, architecture without HyperThreading, limited vector units and cache subsystem when compared to a standard Xeon. Atom has however the advantage that code compiled with standard optimizations (i.e. -O2) runs unmodified on it.
IBM POWER8: The POWER8 is the latest incarnation of the POWER product line which is the evolution of the old PowerPC. While the latter was the result of a strategic alliance between Apple, Motorola and IBM, the former, initially an IBM only brand, is now managed by an industry consortium, Open POWER Foundation , which includes big players like Google, NVidia, Tyan. Compared to the past, the POWER8 includes efforts to simplify platform ports from x86_64 and has the usual focus on highly threaded workloads, by providing an high number of cores.
Intel Xeon Phi: the Xeon Phi is Intel answer to the GPGPU market, which in recent years has dominated the scene of High Performance Computing when the problem to solve maps well on a many-core architecture like the one of GPUs. The Phi consist of a very high number of simplified x86_64 cores, running at a relatively low frequency, which however have a large vector unit. The advantage touted by Intel for such an architecture is that being real x86_64 cores, the porting effort is lower when compared to writing software for a GPU.
3 Test Environments for Power and Performance Measurements
We now describe the test environments we have used to do power and performance measurements for two Intel Xeon processors, belonging respectively to the Sandy Bridge generation and to the Haswell one, an APM ARMv8 64-bit X-Gene1 Server-on-Chip, an Intel Xeon Phi coprocessor, an Intel Atom processor of Avoton class, an IBM POWER8 processor.
3.1 Hardware setup
Table 1 details some of the general details of those processors, in which one can already spot the almost two years advantage which Intel has in terms of fabrication process, when compared to ARM based solutions.
In table 2 actual specifications of the benchmarked models are shown. As one can see from such table there are two obvious tradeoffs which have been made on low power chips (i.e. X-Gene1 and Atom), removing the HyperThreading-like subsystem and keeping the number of cores down. On the other hand the POWER8 went for a much higher number of threads, which is explained by the fact it’s marketed at very high end servers, with highly parallel workloads, e.g. web servers. As already mentioned the Xeon have the ability to scale up single core performance via the so called TurboBoost feature that allows increasing the core frequency when only a few of the cores are being utilized (reported in parenthesis).
|\br||# Cores||# Threads||Frequency (GHz)|
|\mrXeon SandyBridge||8||16||2.0 (2.8)|
|Xeon Haswell||18||36||2.3 (3.6)|
3.2 Benchmarks setup
We selected three benchmarks for our study, which are commonly used in HEP and in particular in CMS to determine the performance of a machine.
ParFullCMS: a standalone Geant4  simulation using a geometry similar to the one used in production by CMS, with simplified physics. This benchmark has been recently adopted across LHC experiments for multithreading studies, since on traditional architectures it already scales to a large number of threads. For our benchmark we used ParFullCMS which ships with Geant4.10.3.
CMSSW RECO: the reconstruction of 100 events with pileup 35 at 25ns, using CMS Offline Software (CMSSW). This is considered a standard candle to measure the performance of CMS reconstruction as it guarantees good coverage of the code associated to all of reconstructed objects. For our benchmark we used the development branch of CMSSW, 7.5.x, as of April 1st 2015.
HEPSPEC06: a HEP driven benchmark suite, widely used as it should scale as HEP specific workloads . In particular it is used for LCG site pledges and therefore purchases.
All the benchmarks were compiled using the latest version of GCC 4.9.x available on a given platform. The only notable exception to this was the Phi benchmarks, which used the Intel Compiler (ICC 15.0.2) given the lack of support for vector units of GCC. Since we wanted to be as close as possible to the production setup used by CMS when running on the grid, we switched on conservative optimization flags (i.e. -O2) and we compiled code targeting vector units, we did not use any platform specific optimization or aggressive optimization option (e.g. -ffast-math). In particular we used the same exact binaries for Xeon and Atom as platform compatibility is considered by Intel one of Atom advantages over other low power architectures in an heterogeneous environment, like the Grid. Of course, this limits the capabilities of the Xeon, which for example supports much advanced vector instructions (AVX and AVX2) compared to Atom (which only supports SSE4) or the other processors, however we deemed that this setup is closer to what is actually done today in the Grid where usually there is no selection of optimized binaries for more recent CPU architecture. Enabling this and benchmarking results is of course interesting, but outside the scope of this paper.
Given Xeon SandyBridge is right now one of the most popular CPUs on the Grid, and in order to simplify comparisons between different benchmarks, we normalized all the results to those of the same benchmark running on in single core mode on our SandyBridge test system.
In the ParFullCMS case we ran a total of 1220 events, subdividing the workload between an increasing number of threads, up to the number of hardware threads the CPU had (in case HyperThreading like mechanism was available) or double the number of hardware cores (in case the CPU did not support hardware). The throughput of the CPU was measured by dividing the number of events by the number of seconds spent in the event processing loop, as reported by ParFullCMS. This number is an underestimation of the performance of a CPU, but it’s deemed to be a good enough approximation, especially for a low number of threads.
In the CMSSW case we started an increasing number of processes, using the same logic as before to determine the maximum number of processes. Each process is running the same amount of events and in the end we calculate the total throughput by summing the throughput of each job. This ends up overestimating the actual throughput of short running jobs, but it’s a closer match to the actual production case.
For the HEPSPEC06 case we started the benchmark specifying the relevant option to run it multithreaded and used the number reported at the end as indication of the performance of the system.
We have previously detailed our methodology on how to measure CPU power utilization [8, 4]. On-chip sensors (Running Average Power Limit – RAPL) were used to measure silicon level power consumption for Xeon CPUs. Similar on-board sensors were available for X-Gene1 and Xeon Phi products. HP Moonshot chassis was equipped with power consumption reporting, but external power distribution unit (PDU) was used instead as it provided data logging capability. Only a single power supply unit (PSU) was used in HP Moonshot systems while conducting experiments. Unfortunately we were not able to get power consumption measurements for the POWER8 CPU or the box. Atom systems did not provide RAPL reporting thus we used a full node power consumption measurements.
5.1 Raw performance
In figure 1 we show the results of all our benchmarks. As anticipated, we can clearly identify two class of performance, Xeons and POWER8 on one side and X-Gene and Atom on the other side.
By looking at figure 3, one can immediately see which CPU has HyperThreading (Xeon, POWER8) and which does not (Atom, X-Gene1). In particular we see how POWER8 hardware threading scales better than the others, thanks to the eight hardware threads per core, but it’s far from perfect scaling when in HyperThreading regime. We attribute the disappointing performance of Xeon Phi to the fact that the benchmark is a direct port of the multithreaded application without any specific Phi improvement and optimization. We nevertheless decided to include the results in this comparison to point out how Xeon Phi (and to a lower degree POWER8) do need a non negligible code optimization effort in order to perform to their maximum.
Similarly in figure 3 we have plotted the per core performance, which immediately highlights how TurboBoost provides Xeon additional performance with a lower CPU usage. This highlights the importance of benchmarking a modern CPU with a load close to the average production one, since single process benchmarks will always overestimate performance.
In all our benchmark the new Intel Xeon is shown to be the best overall performer, both for single thread and fully loaded socket tests. In the fully loaded case, the only contender seems to be the POWER8, with very close HEPSPEC6 absolute results, if one does not consider the different operating frequencies.
5.2 Power efficiency
As we said, raw performance comes with a price tag in terms of power consumption as it can be seen in figure 5. While Haswell performs extremely well, it’s also the one that requires most of the power to run. The actual efficiency of each system can be better evaluated looking at figure 5 which clearly shows that Haswell and Atom are in the same league. While in the past we reported X-Gene1 as a possible contender, it’s also clear that Intel is not sitting idle and without continuous effort to steadily follow an improvement roadmap, newer generation of Intel chips quickly advance not only in terms of raw performance but power efficiency in general. One of the possible metrics to decide how to select between similarly efficient CPUs can be fount in figure 6 which provides power efficiency as a function of performance. Fixing the wanted / required performance level immediately gives the platform which performs better in terms of power efficiency.
5.3 Box to box comparison
For our tests we had the privilege to try out an HP Moonshot chassis, which has the peculiarity of supporting cartridges with Atom (the m300 model) and X-Gene1 (the m400 one). This allowed us to make a box to box comparison where the contribution to the power consumption for the empty chassis is similar and the volume occupied in a rack is exactly the same. In particular in our test setup we had 5 Atom cartridges and 15 X-Gene1 ones. We measured the throughput of ParFullCMS by running it on an increasing number of cartridges, while at the same time measuring the power consumption of the whole box. In figure 7 you have the results for our measurements in form of blue dots and red crosses. Since the two boxes are filled with a different number of cartridges, we cannot compare the two dataset directly, as the power cost due to idle cartridges is different. What we did instead is to used the data we acquired to evaluate the power consumption of a cartridge in running and idle mode and then use those values to project the results for fully populated boxes. The results are extremely unsatisfactory for the X-Gene1 based m400 cartridge which seems to be extremely underperforming in terms of power consumption. We attribute this to actual maturity issues of the cartridge itself, which suspiciously has the same power consumption of the development board, rather than actual power consumption of the CPU. This well illustrates the fact that while promising, a lot needs to be done in term of production readiness for ARMv8 based solution.
We continued our investigation effort in evaluating alternative platforms for HEP workloads. In particular we have extended our analysis to include more benchmarks and novel platforms. While not final, our conclusion shows that depending on the various operating constrains in terms of power usage and performance, either Atom based systems or latest generation Intel Xeon system provide the best power efficiency levels. While APM X-Gene platform is still relevant, it’s clear that in order to compete with Intel, ARM based solutions need not only to develop a performing CPU and match Intel pace in evolving their products, but they need to address the maturity issues of their ecosystems, both in terms of software and of auxiliary electronics. From the mere performance point of view, both the POWER8 and even more so Xeon Phi results show how difficult it is to take advantage of extremely parallel architecture without specifically design software around them.
This work was partially supported by the National Science Foundation, under Cooperative Agreement PHY-1120138, and by the U.S. Department of Energy. We would like to express our gratitude to APM for providing hardware and effort benchmarking Geant4 ParFullCMS, to Intel for providing and managing some of the Intel Xeon used, and to TechLab at CERN for providing and managing other Intel Xeons used, the POWER8 system, the HP Moonshot systems and Intel Xeon Phi server for the benchmarks.
-  Lyndon Evans and Philip Bryant. LHC Machine. JINST, 3:S08001, 2008.
-  S. Chatrchyan et al. The CMS experiment at the CERN LHC. JINST, 3:S08004, 2008.
-  L Rossi and O Bruning. High Luminosity Large Hadron Collider A description for the European Strategy Preparatory Group. Technical Report CERN-ATS-2012-236, CERN, Geneva, Aug 2012.
-  David Abdurachmanov, Peter Elmer, Giulio Eulisse, and Shahzad Muzaffar. Initial explorations of arm processors for scientific computing. Proceedings of ACAT 2014, Prague, 2014.
-  OpenPOWER foundation. http://openpowerfoundation.org.
-  S. Agostinelli et al. GEANT4: A Simulation toolkit. Nucl.Instrum.Meth., A506:250–303, 2003.
-  Michele Michelotto, Manfred Alef, Alejandro Iribarren, Helge Meinhard, Peter Wegner, Martin Bly, Gabriele Benelli, Franco Brasolin, Hubert Degaudenzi, Alessandro De Salvo, Ian Gable, Andreas Hirstius, and Peter Hristov. A comparison of HEP code with SPEC 1 benchmarks on multi-core worker nodes. J. Phys.: Conf. Ser., 219(5):052009, apr 2010.
-  David Abdurachmanov, Peter Elmer, Giulio Eulisse, Robert Knight, Tapio Petteri Niemi, Jukka Nurminen, Filip Nyback, Goncalo Marques Pestana, and Zhonghong Ou. Techniques and tools for measuring energy efficiency of scientific software applications. To be published in ACAT 2014 proceedings, 2014.