Deep Learning Recommendation Models (DLRM) are widespread, account for a considerable data center footprint, and grow by more than 1.5x per year. With model size soon to be in terabytes range, leveraging Storage Class Memory (SCM) for inference enables lower power consumption. This paper evaluates the major challenges in extending the memory hierarchy to SCM for DLRM, and presents different techniques to improve performance through a Software Defined Memory. We show how underlying technologies such as Nand Flash and 3DXP differentiate, and relate to real world scenarios, enabling from 5% to 29% power savings.

Deep learning recommendation models (DLRMs) have been used across many business-critical services at Meta and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper, we present Neo, a software-hardware co-designed system for high-performance distributed training of large-scale DLRMs. Neo employs a novel 4D parallelism strategy that combines table-wise, row-wise, column-wise, and data parallelism for training massive embedding operators in DLRMs. In addition, Neo enables extremely high-performance and memory-efficient embedding computations using a variety of critical systems optimizations, including hybrid kernel fusion, software-managed caching, and quality-preserving compression. Finally, Neo is paired with ZionEX, a new hardware platform co-designed with Neo's 4D parallelism for optimizing communications for large-scale DLRM training. Our evaluation on 128 GPUs using 16 ZionEX nodes shows that Neo outperforms existing systems by up to 40× for training 12-trillion-parameter DLRM models deployed in production.

The widespread application of deep learning has changed the landscape of computation in data centers. In particular, personalized recommendation for content ranking is now largely accomplished using deep neural networks. However, despite their importance and the amount of compute cycles they consume, relatively little research attention has been devoted to recommendation systems. To facilitate research and advance the understanding of these workloads, this paper presents a set of real-world, production-scale DNNs for personalized recommendation coupled with relevant performance metrics for evaluation. In addition to releasing a set of open-source workloads, we conduct in-depth analysis that underpins future system design and optimization for at-scale recommendation: Inference latency varies by 60% across three Intel server generations, batching and co-location of inference jobs can drastically improve latency-bounded throughput, and diversity across recommendation models leads to different optimization strategies.

Personalized recommendation systems leverage deep learning models and account for the majority of data center AI cycles. Their performance is dominated by memory-bound sparse embedding operations with unique irregular memory access patterns that pose a fundamental challenge to accelerate. This paper proposes a lightweight, commodity DRAM compliant, near-memory processing solution to accelerate personalized recommendation inference. The in-depth characterization of production-grade recommendation models shows that embedding operations with high model-, operator and data-level parallelism lead to memory bandwidth saturation, limiting recommendation inference performance. We propose RecNMP which provides a scalable solution to improve system throughput, supporting a broad range of sparse embedding models. RecNMP is specifically tailored to production environments with heavy co-location of operators on a single server. Several hardware/software cooptimization techniques such as memory-side caching, tableaware packet scheduling, and hot entry profiling are studied, providing up to 9.8× memory latency speedup over a highly-optimized baseline. Overall, RecNMP offers 4.2× throughput improvement and 45.8% memory energy savings.

This article consists only of a collection of slides from the author's conference presentation. It is concluded that Zion, FaceBook's next generation flexible training platform, is co-designed to target demanding recommendation models. It adopts new vendor-agnostic OCP accelerator module and it's a Building block that can scale out to a bigger system.

Machine learning sits at the core of many essential products and services at Facebook. This paper describes the hardware and software infrastructure that supports machine learning at global scale. Facebook's machine learning workloads are extremely diverse: services require many different types of models in practice. This diversity has implications at all layers in the system stack. In addition, a sizable fraction of all data stored at Facebook flows through machine learning pipelines, presenting significant challenges in delivering data to high-performance distributed training flows. Computational requirements are also intense, leveraging both GPU and CPU platforms for training and abundant CPU capacity for real-time inference. Addressing these and other emerging challenges continues to require diverse efforts that span machine learning algorithms, software, and hardware design.

The sparse matrix is a key data structure in various domains such as high-performance computing, machine learning, and graph analytics. To maximize performance of sparse matrix operations, it is especially important to optimize across the operations and not just within individual operations. While a straightforward per-operation mapping to library routines misses optimization opportunities, manually optimizing across the boundary of library routines is time-consuming and error-prone, sacrificing productivity.

Stochastic Gradient Descent (SGD) is a popular optimization method used to train a variety of machine learning models. Most of SGD work to-date has concentrated on improving its statistical efficiency, in terms of rate of convergence to the optimal solution. At the same time, as parallelism of modern CPUs continues to increase through progressively higher core counts, it is imperative to understand the parallel hardware efficiency of SGD, which often comes at odds with its statistical efficiency. In this paper, we explore several modern parallelization methods of SGD on a shared memory system, in the context of sparse and convex optimization problems. Specifically, we develop optimized parallel implementations of several SGD algorithms, and show that their parallel efficiency is severely limited by inter-core communication. We propose a new, scalable, communication-avoiding, many-core friendly implementation of SGD, called HogBatch, which exposes parallelism on several levels, minimizes the impact on statistical efficiency, and, as a result significantly outperforms the other methods. On a variety of datasets, HogBatch demonstrates near linear scalability on a system with 14 cores, as well as delivers up to a 20X speedup over previous methods.

As quantum computers of non-trivial size become available in the near future, it is imperative to develop tools to emulate small quantum computers. This allows for validation and debugging of algorithms as well as exploring hardware-software co-design to guide the development of quantum hardware and architectures. The simulation of quantum computers entails multiplications of sparse matrices with very large dense vectors of dimension 2n, where n denotes the number of qubits, making this a memory-bound and network bandwidth-limited application. We introduce the concept of a quantum computer emulator as a component of a software framework for quantum computing, enabling a significant performance advantage over simulators by emulating quantum algorithms at a high level rather than simulating individual gate operations. We describe various optimization approaches and present benchmarking results, establishing the superiority of quantum computer emulators in terms of performance.

BLIS is a new software framework for instantiating high-performance BLAS-like dense linear algebra libraries. We demonstrate how BLIS acts as a productivity multiplier by using it to implement the level-3 BLAS on a variety of current architectures. The systems for which we demonstrate the framework include state-of-the-art general-purpose, low-power, and many-core architectures. We show, with very little effort, how the BLIS framework yields sequential and parallel implementations that are competitive with the performance of ATLAS, OpenBLAS (an effort to maintain and extend the GotoBLAS), and commercial vendor implementations such as AMD’s ACML, IBM’s ESSL, and Intel’s MKL libraries. Although most of this article focuses on single-core implementation, we also provide compelling results that suggest the framework’s leverage extends to the multithreaded domain.

In this work, we revisit the 1999 Gordon Bell Prize winning PETSc-FUN3D aerodynamics code, extending it with highly-tuned shared-memory parallelization and detailed performance analysis on modern highly parallel architectures. An unstructured-grid implicit flow solver, which forms the backbone of computational aerodynamics, poses particular challenges due to its large irregular working sets, unstructured memory accesses, and variable/limited amount of parallelism. This code, based on a domain decomposition approach, exposes tradeoffs between the number of threads assigned to each MPI-rank sub domain, and the total number of domains. By applying several algorithm- and architecture-aware optimization techniques for unstructured grids, we show a 6.9X speed-up in performance on a single-node Intel® XeonTM1 E5 2690 v2 processor relative to the out-of-the-box compilation. Our scaling studies on TACC Stampede supercomputer show that our optimizations continue to provide performance benefits over baseline implementation as we scale up to 256 nodes.

Algebraic Multigrid (AMG) is a linear solver, well known for its linear computational complexity and excellent parallelization scalability. As a result, AMG is expected to be a solver of choice for emerging extreme scale systems capable of delivering hundred Pflops and beyond. While node level performance of AMG is generally limited by memory bandwidth, achieving high bandwidth efficiency is challenging due to highly sparse irregular computation, such as triple sparse matrix products, sparse-matrix dense-vector multiplications, independent set coarsening algorithms, and smoothers such as Gauss-Seidel. We develop and analyze a highly optimized AMG implementation, based on the well-known HYPRE library. Compared to the HYPRE baseline implementation, our optimized implementation achieves 2.0x speedup on a recent Intel® Xeon® Haswell processor. Combined with our other multi-node optimizations, this translates into similarly high speedups when weak-scaled multiple nodes. In addition, our implementation achieves 1.3x speedup compared to AmgX, NVIDIA's high-performance implementation of AMG, running on K40c.

BLIS is a new framework for rapid instantiation of the BLAS. We describe how BLIS extends the "GotoBLAS approach" to implementing matrix multiplication (GEMM). While GEMM was previously implemented as three loops around an inner kernel, BLIS exposes two additional loops within that inner kernel, casting the computation in terms of the BLIS micro-kernel so that porting G E M M becomes a matter of customizing this micro-kernel for a given architecture. We discuss how this facilitates a finer level of parallelism that greatly simplifies the multithreading of GEMM as well as additional opportunities for parallelizing multiple loops. Specifically, we show that with the advent of many-core architectures such as the IBM PowerPC A2 processor (used by Blue Gene/Q) and the Intel Xeon Phi processor, parallelizing both within and around the inner kernel, as the BLIS approach supports, is not only convenient, but also necessary for scalability. The resulting implementations deliver what we believe to be the best open source performance for these architectures, achieving both impressive performance and excellent scalability.

Intel Xeon Phi coprocessor-based clusters offer high compute and memory performance for parallel workloads and also support direct network access. Many real world applications are significantly impacted by network characteristics and to maximize the performance of such applications on these clusters, it is particularly important to effectively saturate network bandwidth and/or hide communications latency. We demonstrate how to do so using techniques such as pipelined DMAs for data transfer, dynamic chunk sizing, and better asynchronous progress. We also show a method for, and the impact of avoiding serialization and maximizing parallelism during application communication phases. Additionally, we apply application optimizations focused on balancing computation and communication in order to hide communication latency and improve utilization of cores and of network bandwidth. We demonstrate the impact of our techniques on three well known and highly optimized HPC kernels running natively on the Intel Xeon Phi coprocessor. For the Wilson-Dslash operator from Lattice QCD, we characterize the improvements from each of our optimizations for communication performance, apply our method for maximizing concurrency during communication phases, and show an overall 48% improvement from our previously best published result. For HPL/LINPACK, we show 68.5% efficiency with 97 TFLOPs on 128 Intel Xeon Phi coprocessors, the first ever reported native HPL efficiency on a coprocessor-based supercomputer. For FFT, we show 10.8 TFLOPs using 1024 Intel Xeon Phi coprocessors on the TACC Stampede cluster, the highest reported performance on any Intel Architecture-based cluster and the first such result to be reported on a coprocessor-based supercomputer.

We present an end-to-end optimization of the innovative Arbitrary high-order DERivative Discontinuous Galerkin (ADER-DG) software SeisSol targeting Intel® Xeon Phi coprocessor platforms, achieving unprecedented earthquake model complexity through coupled simulation of full frictional sliding and seismic wave propagation. SeisSol exploits unstructured meshes to flexibly adapt for complicated geometries in realistic geological models. Seismic wave propagation is solved simultaneously with earthquake faulting in a multiphysical manner leading to a heterogeneous solver structure. Our architecture aware optimizations deliver up to 50% of peak performance, and introduce an efficient compute-communication overlapping scheme shadowing the multiphysics computations. SeisSol delivers near-optimal weak scaling, reaching 8.6 DP-PFLOPS on 8,192 nodes of the Tianhe-2 supercomputer. Our performance model projects reaching 18 -- 20 DP-PFLOPS on the full Tianhe-2 machine. Of special relevance to modern civil engineering needs, our pioneering simulation of the 1992 Landers earthquake shows highly detailed rupture evolution and ground motion at frequencies up to 10 Hz.

The gap between the cost of moving data and the cost of computing continues to grow, making it ever harder to design iterative solvers on extreme-scale architectures. This problem can be alleviated by alternative algorithms that reduce the amount of data movement. We investigate this in the context of Lattice Quantum Chromo dynamics and implement such an alternative solver algorithm, based on domain decomposition, on Intel®Xeon Phi co-processor (KNC) clusters. We demonstrate close-to-linear on-chip scaling to all 60 cores of the KNC. With a mix of single- and half-precision the domain-decomposition method sustains 400-500 Gflop/s per chip. Compared to an optimized KNC implementation of a standard solver [1], our full multi-node domain-decomposition solver strong-scales to more nodes and reduces the time-to-solution by a factor of 5.

A new sparse high performance conjugate gradient benchmark (HPCG) has been recently released to address challenges in the design of sparse linear solvers for the next generation extreme-scale computing systems. Key computation, data access, and communication pattern in HPCG represent building blocks commonly found in today's HPC applications. While it is a well known challenge to efficiently parallelize Gauss-Seidel smoother, the most time-consuming kernel in HPCG, our algorithmic and architecture-aware optimizations deliver 95% and 68% of the achievable bandwidth on Xeon and Xeon Phi, respectively. Based on available parallelism, our Xeon Phi shared-memory implementation of Gauss-Seidel smoother selectively applies block multi-color reordering. Combined with MPI parallelization, our implementation balances parallelism, data access locality, CG convergence rate, and communication overhead. Our implementation achieved 580 TFLOPS (82% parallelization efficiency) on Tianhe-2 system, ranking first on the most recent HPCG list in July 2014. In addition, we demonstrate that our optimizations not only benefit HPCG original dataset, which is based on structured 3D grid, but also a wide range of unstructured matrices.

Sparse matrix-vector multiplication (SpMV) is an important kernel in many scientific applications and is known to be memory bandwidth limited. On modern processors with wide SIMD and large numbers of cores, we identify and address several bottlenecks which may limit performance even before memory bandwidth: (a) low SIMD efficiency due to sparsity, (b) overhead due to irregular memory accesses, and (c) load-imbalance due to non-uniform matrix structures.

Dense linear algebra has been traditionally used to evaluate the performance and efficiency of new architectures. This trend has continued for the past half decade with the advent of multi-core processors and hardware accelerators. In this paper we describe how several flavors of the Linpack benchmark are accelerated on Intel's recently released Intel®Xeon Phi™1co-processor (code-named Knights Corner) in both native and hybrid configurations. Our native DGEMM implementation takes full advantage of Knights Corner's salient architectural features and successfully utilizes close to 90% of its peak compute capability. Our native Linpack implementation running entirely on Knights Corner employs novel dynamic scheduling and achieves close to 80% efficiency - the highest published co-processor efficiency. Similarly to native, our single-node hybrid implementation of Linpack also achieves nearly 80% efficiency. Using dynamic scheduling and an enhanced look-ahead scheme, this implementation scales well to a 100-node cluster, on which it achieves over 76% efficiency while delivering the total performance of 107 TFLOPS.

We analyse gather-scatter performance bottlenecks in molecular dynamics codes and the challenges that they pose for obtaining benefits from SIMD execution. This analysis informs a number of novel code-level and algorithmic improvements to Sandia's miniMD benchmark, which we demonstrate using three SIMD widths (128-, 256and 512bit). The applicability of these optimisations to wider SIMD is discussed, and we show that the conventional approach of exposing more parallelism through redundant computation is not necessarily best. In single precision, our optimised implementation is up to 5x faster than the original scalar code running on Intel®Xeon®processors with 256-bit SIMD, and adding a single Intel®Xeon Phi™coprocessor provides up to an additional 2x performance increase. These results demonstrate: (i) the importance of effective SIMD utilisation for molecular dynamics codes on current and future hardware; and (ii) the considerable performance increase afforded by the use of Intel®Xeon Phi™coprocessors for highly parallel workloads.

The Non-Uniform Fast Fourier Transform (NUFFT) is a generalization of FFT to non-equidistant samples. It has many applications which vary from medical imaging to radio astronomy to the numerical solution of partial differential equations. Despite recent advances in speeding up NUFFT on various platforms, its practical applications are still limited, due to its high computational cost, which is significantly dominated by the convolution of a signal between a non-uniform and uniform grids. The computational cost of the NUFFT is particularly detrimental in cases which require fast reconstruction times, such as iterative 3D non-Cartesian MRI reconstruction. We propose novel and highly scalable parallel algorithm for performing NUFFT on x86-based multi-core CPUs. The high performance of our algorithm relies on good SIMD utilization and high parallel efficiency. On convolution, we demonstrate on average 90% SIMD efficiency using SSE, as well up to linear scalability using a quad-socket 40-core Intel(R) Xeon(R) E7-4870 Processors based system. As a result, on dual socket Intel(R) Xeon(R) X5670 based server, our NUFFT implementation is more than 4x faster compared to the best available NUFFT3D implementation, when run on the same hardware. On Intel(R) Xeon(R) E5-2670 processor based server, our NUFFT implementation is 1.5X faster than any published NUFFT implementation today. Such speed improvement opens new usages for NUFFT. For example, iterative multi channel reconstruction of a 240×240×240 image could execute in just over 3 minutes, which is on the same order as contemporary non-iterative (and thus less-accurate) 3D NUFFT-based MRI reconstructions.

This paper presents an algorithmic approach for improving the performance of many types of stochastic dynamical simulations. The approach is to redesign existing algorithms that use sparse matrix-vector products (SPMV) with single vectors to instead use a more efficient kernel, the generalized SPMV (GSPMV), which computes with multiple vectors simultaneously. In this paper, we show how to redesign a dynamical simulation to exploit GSPMV in way that is not initially obvious because only one vector is available at a time. We study the performance of GSPMV as a function of the number of vectors, and demonstrate the use of GSPMV in the Stokesian dynamics method for the simulation of the motion of macromolecules in the cell. Specifically, for our application, we find that with modern multicore Intel microprocessors in clusters of up to 64 nodes, we can typically multiply by 8 to 16 vectors in only twice the time required to multiply by a single vector. After redesigning the Stokesian dynamics algorithm to exploit GSPMV, we measure a 30 percent speedup in performance in single-node, data parallel simulations.

Current processor trends of integrating more cores with wider SIMD units, along with a deeper and complex memory hierarchy, have made it increasingly more challenging to extract performance from applications. It is believed by some that traditional approaches to programming do not apply to these modern processors and hence radical new languages must be discovered. In this paper, we question this thinking and offer evidence in support of traditional programming methods and the performance-vs-programming effort effectiveness of common multi-core processors and upcoming manycore architectures in delivering significant speedup, and close-to-optimal performance for commonly used parallel computing workloads. We first quantify the extent of the “Ninja gap”, which is the performance gap between naively written C/C++ code that is parallelism unaware (often serial) and best-optimized code on modern multi-/many-core processors. Using a set of representative throughput computing benchmarks, we show that there is an average Ninja gap of 24X (up to 53X) for a recent 6-core Intel®Core™ i7 X980 Westmere CPU, and that this gap if left unaddressed will inevitably increase. We show how a set of well-known algorithmic changes coupled with advancements in modern compiler technology can bring down the Ninja gap to an average of just 1.3X. These changes typically require low programming effort, as compared to the very high effort in producing Ninja code. We also discuss hardware support for programmability that can reduce the impact of these changes and even further increase programmer productivity. We show equally encouraging results for the upcoming Intel®Many Integrated Core architecture (Intel®MIC) which has more cores and wider SIMD. We thus demonstrate that we can contain the otherwise uncontrolled growth of the Ninja gap and offer a more stable and predictable performance growth over future architectures, offering strong evidence that radical language changes are not required.

Tackling computationally challenging problems with high efficiency often requires the combination of algorithmic innovation, advanced architecture, and thorough exploitation of parallelism. We demonstrate this synergy through synthetic aperture radar (SAR) via backprojection, an image reconstruction method that can require hundreds of TFLOPS. Computation cost is significantly reduced by our new algorithm of approximate strength reduction; data movement cost is economized by software locality optimizations facilitated by advanced architecture support; parallelism is fully harnessed in various patterns and granularities. We deliver over 35 billion backprojections per second throughput per compute node on an Intel®Xeon®processor E5-2670-based cluster, equipped with Intel®Xeon Phi coprocessors. This corresponds to processing a 3K×3K image within a second using a single node. Our study can be extended to other settings: backprojection is applicable elsewhere including medical imaging, approximate strength reduction is a general code transformation technique, and many-core processors are emerging as a solution to energy-efficient computing.

Multigrid methods are widely used to accelerate the convergence of iterative solvers for linear systems used in a number of different application areas. In this paper, we explore optimization techniques for geometric multigrid on existing and emerging multicore systems including the Opteron-based Cray XE6, Intel® Xeon® E5-2670 and X5550 processor-based Infiniband clusters, as well as the new Intel® Xeon Phi coprocessor (Knights Corner). Our work examines a variety of novel techniques including communication-aggregation, threaded wavefront-based DRAM communication-avoiding, dynamic threading decisions, SIMDization, and fusion of operators. We quantify performance through each phase of the V-cycle for both single-node and distributed-memory experiments and provide detailed analysis for each class of optimization. Results show our optimizations yield significant speedups across a variety of subdomain sizes while simultaneously demonstrating the potential of multi- and manycore processors to dramatically accelerate single-node performance. However, our analysis also indicates that improvements in networks and communication will be essential to reap the potential of manycore processors in large-scale multigrid calculations.

Lattice Quantum Chromo-dynamics (LQCD) is a computationally challenging problem that solves the discretized Dirac equation in the presence of an SU(3) gauge field. Its key operation is a matrix-vector product, known as the Dslash operator. We have developed a novel multicore architecture-friendly implementation of the Wilson-Dslash operator which delivers 75 Gflops (single-precision) on an Intel® Xeon® Processor X5680 achieving 60% computational efficiency for datasets that fit in the last-level cache. For datasets larger than the last-level cache, this performance drops to 50 Gflops. Our performance is 2-3X higher than a well-known implementation from the Chroma software suite when running on the same hardware platform. The novel implementation of LQCD reported in this paper is based on recently published the 3.5D spatial and 4.5D temporal tiling schemes. Both blocking schemes significantly reduce LQCD external memory bandwidth requirements, delivering a more compute-bound implementation. The performance advantage of our schemes will become more significant as the gap between compute flops and external memory bandwidth continues to grow. We demonstrate very good cluster-level scalability of our implementation: for a lattice of 323 x 256 sites, we achieve over 4 Tflops when strong-scaled to a 128 node system (1536 cores total). For the same lattice size, a full Conjugate Gradients Wilson-Dslash operator, achieves 2.95 Tflops.

Recent advances in computing have led to an explosion in the amount of data being generated. Processing the ever-growing data in a timely manner has made throughput computing an important aspect for emerging applications. Our analysis of a set of important throughput computing kernels shows that there is an ample amount of parallelism in these kernels which makes them suitable for today's multi-core CPUs and GPUs. In the past few years there have been many studies claiming GPUs deliver substantial speedups (between 10X and 1000X) over multi-core CPUs on these kernels. To understand where such large performance difference comes from, we perform a rigorous performance analysis and find that after applying optimizations appropriate for both CPUs and GPUs the performance gap between an Nvidia GTX280 processor and the Intel Core i7-960 processor narrows to only 2.5x on average. In this paper, we discuss optimization techniques for both CPU and GPU, analyze what architecture features contributed to performance differences between the two architectures, and recommend a set of architectural features which provide significant improvement in architectural efficiency for throughput kernels.

Medical volumetric imaging requires high fidelity, high performance rendering algorithms. We motivate and analyze new volumetric rendering algorithms that are suited to modern parallel processing architectures. First, we describe the three major categories of volume rendering algorithms and confirm through an imaging scientist-guided evaluation that ray-casting is the most acceptable. We describe a thread- and data-parallel implementation of ray-casting that makes it amenable to key architectural trends of three modern commodity parallel architectures: multi-core, GPU, and an upcoming many-core Intel architecture code-named Larrabee. We achieve more than an order of magnitude performance improvement on a number of large 3D medical datasets. We further describe a data compression scheme that significantly reduces data-transfer overhead. This allows our approach to scale well to large numbers of Larrabee cores.

The current trend is for processors to deliver dramatic improvements in parallel performance while only modestly improving serial performance. Parallel performance is harvested through vector/SIMD instructions as well as multithreading (through both multithreaded cores and chip multiprocessors). Vector parallelism can be more efficiently supported than multithreading, but is often harder for software to exploit. In particular, code with sparse data access patterns cannot easily utilize the vector/SIMD instructions of mainstream processors. Hardware to scatter and gather sparse data has previously been proposed to enable vector execution for these codes. However, on multithreaded architectures, a number of applications spend significant time on atomic operations (e.g., parallel reductions), which cannot be vectorized using previously proposed schemes. This paper proposes architectural support for atomic vector operations (referred to as GLSC) that addresses this limitation. GLSC extends scatter-gather hardware to support atomic memory operations. Our experiments show that the GLSC provides an average performance improvement on a set of important RMS kernels of 54% for 4-wide SIMD.

This paper examines the growing need for a general-purpose ldquoanalytics enginerdquo that can enable next-generation processing platforms to effectively model events, objects, and concepts based on end-user input, and accessible datasets, along with an ability to iteratively refine the model in real-time. We find such processing needs at the heart of many emerging applications and services. This processing is further decomposed in terms of an integration of three fundamental compute capabilities-recognition, mining, and synthesis (RMS). The set of RMS workloads is examined next in terms of usage, mathematical models, numerical algorithms, and underlying data structures. Our analysis suggests a workload convergence that is analyzed next for its platform implications. In summary, a diverse set of emerging RMS applications from market segments like graphics, gaming, media-mining, unstructured information management, financial analytics, and interactive virtual communities presents a relatively focused, highly overlapping set of common platform challenges. A general-purpose processing platform designed to address these challenges has the potential for significantly enhancing users' experience and programmer productivity.

In this paper we describe parallelization of interior-point method (IPM) aimed at achieving high scalability on large-scale chip-multiprocessors (CMPs). IPM is an important computational technique used to solve optimization problems in many areas of science, engineering and finance. IPM spends most of its computation time in a few sparse linear algebra kernels. While each of these kernels contains a large amount of parallelism, sparse irregular datasets seen in many optimization problems make parallelism difficult to exploit. As a result, most researchers have shown only a relatively low scalability of 4X-12X on medium to large scale parallel machines.

Predicated execution enables the removal of branches by converting segments of branching code into sequences of conditional operations. An important side effect of this transformation is that the compiler must unconditionally assign resources to predicated operations. However, a resource is only put to productive use when the predicate associated with an operation evaluates to True. To reduce this superfluous commitment of resources, we propose probabilistic predicate-aware scheduling to assign multiple operations to the same resource at the same time, thereby over-subscribing its use. Assignment is performed in a probabilistic manner using a combination of predicate profile information and predicate analysis aimed at maximizing the benefits of over-subscription in view of the expected degree of conflict. Conflicts occur when two or more operations assigned to the same resource have their predicates evaluate to True. A predicate-aware VLIW processor pipeline detects such conflicts, recovers, and correctly executes the conflicting operations. By increasing the effective throughput of a fixed set of resources, probabilistic predicate-aware scheduling provided an average of 20% performance gain in our evaluations on a 4-issue processor, and 8% gain on a 6-issue processor.

Register bypass provides additional datapaths to eliminate data hazards in processor pipelines. The difficulty with register bypass is that the cost of the bypass network is substantial and grows substantially as processor width or pipeline depth are increased. For a single application, many of the bypass paths have extremely low utilization. Thus, there is an important opportunity in the design of application-specific processors to remove a large fraction of the bypass cost while maintaining performance comparable to a processor with full bypass. We propose a systematic design customization process along with a bypass-cognizant compiler scheduler. For the former, we employ iterative design space exploration wherein successive processor designs are selected based on bypass utilization statistics combined with the availability of redundant bypass paths. Compiler scheduling for sparse bypass processors is accomplished by prioritizing function unit choices for each operation prior to scheduling using global information. Results show that for a 5-issue customized VLIW processor, 70% of the bypass cost is eliminated while sacrificing only 10% performance.

Predicated execution enables the removal of branches wherein segments of branching code are converted into straight-line segments of conditional operations. An important, but generally ignored side effect of this transformation is that the compiler must assign distinct resources to all the predicated operations at a given time to ensure that those resources are available at run-time. However, a resource is only put to productive use when the predicates associated with its operations evaluate to True. We propose predicate-aware scheduling to reduce the superfluous commitment of resources to operations whose predicates evaluate to False at run-time. The central idea is to assign multiple operations to the same resource at the same time, thereby oversubscribing its use. This assignment is intelligently performed to ensure that no two operations simultaneously assigned to the same resource will have both of their predicates evaluate to True. Thus, no resource is dynamically oversubscribed. The overall effect of predicate aware scheduling is to use resources more efficiently, thereby increasing performance when resource constraints are a bottleneck.

As processor performance increases, there is a corresponding increase in the demands on the memory system, including caches. Research papers have proposed partitioning the cache into instruction/data, temporal/non-temporal, and/or stack/non-stack regions. Each of these designs can improve performance by constructing two separate structures which can be probed in parallel while reducing contention. In this paper, we propose a new memory organization that partitions data references into stack and nonstack regions. Non-stack references are routed to a conventional cache. Stack references, on the other hand, are shown to have several characteristics that can be leveraged to improve performance using a less conventional storage organization. This paper enumerates those characteristics and proposes a new microarchitectural feature, the stack value file (SVF), which exploits them to improve instruction-level parallelism, reduce stack access latencies, reduce demand on the first-level cache, and reduce data bus traffic. Our results show that the SVF can improve execution performance by 29 to 65% while reducing overhead traffic for the stack region by many orders of magnitude over cache structures of the same size.

In this paper, we examine the effectiveness of a new hardware mechanism, called register queues (RQs), which effectively decouples the architected register space from the physical registers. Using RQs, the compiler can allocate physical registers to store live values in the software pipelined loop while minimizing the pressure placed on architected registers. We show that decoupling the architected register space from the physical register space can greatly increase the applicability of software pipelining, even as memory latencies increase. RQs combine the major aspects of existing rotating register file and register connection techniques to generate efficient software pipeline schedules. Through the use of RQs, we can minimize the register pressure and code expansion caused by software pipelining. We demonstrate the effect of incorporating register queues and software pipelining with 983 loops taken from the Perfect Club, the SPEC suites, and the Livermore Kernels.