A single-ISA heterogeneous multi-core architecture is achip multiprocessor composed of cores of varying size, performance,and complexity. This paper demonstrates that thisarchitecture can provide significantly higher performance inthe same area than a conventional chip multiprocessor. It doesso by matching the various jobs of a diverse workload to thevarious cores. This type of architecture covers a spectrum ofworkloads particularly well, providing high single-thread performancewhen thread parallelism is low, and high throughputwhen thread parallelism is high.This paper examines two such architectures in detail,demonstrating dynamic core assignment policies that providesignificant performance gains over naive assignment, andeven outperform the best static assignment. It examines policiesfor heterogeneous architectures both with and withoutmultithreading cores. One heterogeneous architecture we examineoutperforms the comparable-area homogeneous architectureby up to 63%, and our best core assignment strategyachieves up to 31% speedup over a naive policy.

/pdf/single-isa-heterogeneous-multi-core-architectures-for-1hsu8f3e08.pdf

Single-ISA Heterogeneous Multi-Core Architectures for Multithreaded Workload Performance

This paper describes the polymorphous TRIPS architecture which can be configured for different granularities and types of parallelism. TRIPS contains mechanisms that enable the processing cores and the on-chip memory system to be configured and combined in different modes for instruction, data, or thread-level parallelism. To adapt to small and large-grain concurrency, the TRIPS architecture contains four out-of-order, 16-wide-issue Grid Processor cores, which can be partitioned when easily extractable fine-grained parallelism exists. This approach to polymorphism provides better performance across a wide range of application types than an approach in which many small processors are aggregated to run workloads with irregular parallelism. Our results show that high performance can be obtained in each of the three modes--ILP, TLP, and DLP-demonstrating the viability of the polymorphous coarse-grained approach for future microprocessors.

/pdf/exploiting-ilp-tlp-and-dlp-with-the-polymorphous-trips-8yihgkqe4d.pdf

Exploiting ILP, TLP, and DLP with the polymorphous TRIPS architecture

This paper evaluates the Raw microprocessor. Raw addresses thechallenge of building a general-purpose architecture that performswell on a larger class of stream and embedded computing applicationsthan existing microprocessors, while still running existingILP-based sequential programs with reasonable performance in theface of increasing wire delays. Raw approaches this challenge byimplementing plenty of on-chip resources - including logic, wires,and pins - in a tiled arrangement, and exposing them through a newISA, so that the software can take advantage of these resources forparallel applications. Raw supports both ILP and streams by routingoperands between architecturally-exposed functional units overa point-to-point scalar operand network. This network offers lowlatency for scalar data transport. Raw manages the effect of wiredelays by exposing the interconnect and using software to orchestrateboth scalar and stream data transport.We have implemented a prototype Raw microprocessor in IBM's180 nm, 6-layer copper, CMOS 7SF standard-cell ASIC process. Wehave also implemented ILP and stream compilers. Our evaluationattempts to determine the extent to which Raw succeeds in meetingits goal of serving as a more versatile, general-purpose processor.Central to achieving this goal is Raw's ability to exploit all formsof parallelism, including ILP, DLP, TLP, and Stream parallelism.Specifically, we evaluate the performance of Raw on a diverse setof codes including traditional sequential programs, streaming applications,server workloads and bit-level embedded computation.Our experimental methodology makes use of a cycle-accurate simulatorvalidated against our real hardware. Compared to a 180 nmPentium-III, using commodity PC memory system components, Rawperforms within a factor of 2x for sequential applications with a verylow degree of ILP, about 2x to 9x better for higher levels of ILP, and10x-100x better when highly parallel applications are coded in astream language or optimized by hand. The paper also proposes anew versatility metric and uses it to discuss the generality of Raw.

/pdf/evaluation-of-the-raw-microprocessor-an-exposed-wire-delay-xbkh9dz129.pdf

Evaluation of the Raw Microprocessor: An Exposed-Wire-Delay Architecture for ILP and Streams

SIMD organizations amortize the area and power of fetch, decode, and issue logic across multiple processing units in order to maximize throughput for a given area and power budget. However, throughput is reduced when a set of threads operating in lockstep (a warp) are stalled due to long latency memory accesses. The resulting idle cycles are extremely costly. Multi-threading can hide latencies by interleaving the execution of multiple warps, but deep multi-threading using many warps dramatically increases the cost of the register files (multi-threading depth x SIMD width), and cache contention can make performance worse. Instead, intra-warp latency hiding should first be exploited. This allows threads that are ready but stalled by SIMD restrictions to use these idle cycles and reduces the need for multi-threading among warps. This paper introduces dynamic warp subdivision (DWS), which allows a single warp to occupy more than one slot in the scheduler without requiring extra register file space. Independent scheduling entities allow divergent branch paths to interleave their execution, and allow threads that hit to run ahead. The result is improved latency hiding and memory level parallelism (MLP). We evaluate the technique on a coherent cache hierarchy with private L1 caches and a shared L2 cache. With an area overhead of less than 1%, experiments with eight data-parallel benchmarks show our technique improves performance on average by 1.7X.

/pdf/dynamic-warp-subdivision-for-integrated-branch-and-memory-2urw2m0zsl.pdf

Dynamic warp subdivision for integrated branch and memory divergence tolerance

Most implementations of the Single Instruction Multiple Data (SIMD) model available today require that data elements be packed in vector registers. Operations on disjoint vector elements are not supported directly and require explicit data reorganization manipulations. Computations on non-contiguous and especially interleaved data appear in important applications, which can greatly benefit from SIMD instructions once the data is reorganized properly. Vectorizing such computations efficiently is therefore an ambitious challenge for both programmers and vectorizing compilers. We demonstrate an automatic compilation scheme that supports effective vectorization in the presence of interleaved data with constant strides that are powers of 2, facilitating data reorganization. We demonstrate how our vectorization scheme applies to dominant SIMD architectures, and present experimental results on a wide range of key kernels, showing speedups in execution time up to 3.7 for interleaving levels (stride) as high as 8.

/pdf/auto-vectorization-of-interleaved-data-for-simd-4sbafamzk9.pdf

Auto-vectorization of interleaved data for SIMD

Tarantula is an aggressive floating point machine targeted at technical, scientific and bioinformatics workloads, originally planned as a follow-on candidate to the EV8 processor [6, 5]. Tarantula adds to the EV8 core a vector unit capable of 32 double-precision flops per cycle. The vector unit fetches data directly from a 16 MByte second level cache with a peak bandwidth of sixty four 64-bit values per cycle. The whole chip is backed by a memory controller capable of delivering over 64 GBytes/s of raw band- width. Tarantula extends the Alpha ISA with new vector instructions that operate on new architectural state. Salient features of the architecture and implementation are: (1) it fully integrates into a virtual-memory cache-coherent system without changes to its coherency protocol, (2) provides high bandwidth for non-unit stride memory accesses, (3) supports gather/scatter instructions efficiently, (4) fully integrates with the EV8 core with a narrow, streamlined interface, rather than acting as a co-processor, (5) can achieve a peak of 104 operations per cycle, and (6) achieves excellent "real-computation" per transistor and per watt ratios. Our detailed simulations show that Tarantula achieves an average speedup of 5X over EV8, out of a peak speedup in terms of flops of 8X. Furthermore, performance on gather/scatter intensive benchmarks such as Radix Sort is also remarkable: a speedup of almost 3X over EV8 and 15 sustained operations per cycle. Several benchmarks exceed 20 operations per cycle.

/pdf/tarantula-a-vector-extension-to-the-alpha-architecture-2tpkrhr2uy.pdf

Tarantula: a vector extension to the alpha architecture

In one embodiment, the present invention includes an apparatus having a register file to store vector data, an address generator coupled to the register file to generate addresses for a vector memory operation, and a controller to generate an output slice from one or more slices each including multiple addresses, where the output slice includes addresses each corresponding to a separately addressable portion of a memory. Other embodiments are described and claimed.

Implementing vector memory operations

The memory subsystem is responsible for a large fraction of the energy consumed by compute nodes in High Performance Computing (HPC) systems. The rapid increase in the number of cores has been accompanied by a corresponding increase in the DRAM capacity and bandwidth, and as a result, the memory system consumes a significant amount of the power budget available to a compute node. Consequently, there is a broad research effort focused on power management techniques using DRAM low-power modes. However, memory power management continues to present many challenges. In this paper, we study the potential of Dynamic Voltage and Frequency Scaling (DVFS) of the memory subsystems, and consider the ability to select different frequencies for different memory channels. Our approach is based on tuning voltage and frequency dynamically to maximize the energy savings while maintaining performance degradation within tolerable limits. We assume that HPC applications do not demand maximum bandwidth throughout the entire period of execution. We can use these low memory demand intervals to tune down the frequency and, as a result, applications can tolerate a reduction in bandwidth to save energy. In this paper, we study application channel access patterns, and use these patterns to determine potential additional energy savings that can be achieved by accordingly controlling the channels independently. We then evaluate the proposed DVFS algorithm using a novel hybrid evaluation methodology that includes simulation as well as executions on real hardware. Our results demonstrate the large potential of adaptive memory power management techniques based on DVFS for HPC workloads.

/pdf/adaptive-memory-power-management-techniques-for-hpc-soypfiq9qm.pdf

Isaac Hernandez

Papers

Tarantula: a vector extension to the alpha architecture

Implementing vector memory operations

Adaptive memory power management techniques for HPC workloads