Efficient Dynamic Automatic Memory Management And Concurrent Kernel Execution For General-Purpose Programs On Graphics Processing Units

Abstract

Modern supercomputers now use accelerators to achieve their performance with the most widely used accelerator being the Graphics Processing Unit (GPU). However, achieving the performance potential of systems that combine a GPU and CPU is an arduous task which could be made easier with the assistance of the compiler or runtime. In particular, exploiting two features of GPU architectures -- distributed memory and concurrent kernel execution -- is critical to achieve good performance, but in current GPU programming systems, programmers must exploit them manually. This can lead to poor performance. In this thesis, we propose automatic techniques that: i) perform data transfers between the CPU and GPU, ii) allocate resources for concurrent kernels, and iii) schedule concurrent kernels efficiently without programmer intervention.

<p>Most GPU programs access data in GPU memory for performance. Manually inserting data transfers that move data to and from this GPU memory is an error-prone and tedious task. In this work, we develop a software coherence mechanism to fully automate all data transfers between the CPU and GPU without any assistance from the programmer. Our mechanism uses compiler analysis to identify potential stale data accesses and uses a runtime to initiate transfers as necessary. This avoids redundant transfers that are exhibited by all other existing automatic memory management proposals for general purpose programs. We integrate our automatic memory manager into the X10 compiler and runtime, and find that it not only results in smaller and simpler programs, but also eliminates redundant memory transfers. Tested on eight programs ported from the Rodinia benchmark suite it achieves (i) a 1.06x speedup over hand-tuned manual memory management, and (ii) a 1.29x speedup over another recently proposed compiler--runtime automatic memory management system. Compared to other existing runtime-only (ADSM) and compiler-only (OpenMPC) proposals, it also transfers 2.2x to 13.3x less data on average.

<p>Each new generation of GPUs vastly increases the resources available to GPGPU programs. GPU programming models (like CUDA) were designed to scale to use these resources. However, we find that CUDA programs actually do not scale to utilize all available resources, with over 30% of resources going unused on average for programs of the Parboil2 suite. Current GPUs therefore allow concurrent execution of kernels to improve utilization. We study concurrent execution of GPU kernels using multiprogrammed workloads on current NVIDIA Fermi GPUs. On two-program workloads from Parboil2 we find concurrent execution is often no better than serialized execution. We identify lack of control over resource allocation to kernels as a major serialization bottleneck. We propose transformations that convert CUDA kernels into elastic kernels which permit fine-grained control over their resource usage. We then propose several elastic-kernel aware runtime concurrency policies that offer significantly better performance and concurrency than the current CUDA policy. We evaluate our proposals on real hardware using multiprogrammed workloads constructed from benchmarks in the Parboil2 suite. On average, our proposals increase system throughput (STP) by 1.21x and improve the average normalized turnaround time (ANTT) by 3.73x for two-program workloads over the current CUDA concurrency implementation.

<p>Recent NVIDIA GPUs use a FIFO policy in their thread block scheduler (TBS) to schedule thread blocks of concurrent kernels. We show that FIFO leaves performance to chance, resulting in significant loss of performance and fairness. To improve performance and fairness, we propose use of the Shortest Remaining Time First (SRTF) policy instead. Since SRTF requires an estimate of runtime (i.e. execution time), we introduce Structural Runtime Prediction that uses the grid structure of GPU programs for predicting runtimes. Using a novel Staircase model of GPU kernel execution, we show that kernel runtime can be predicted by profiling only the first few thread blocks. We evaluate an online predictor based on this model on benchmarks from ERCBench and find that predictions made after the execution of single thread block are between 0.48x to 1.08x of actual runtime. %Next, we design a thread block scheduler that is both concurrent kernel-aware and incorporates this predictor. We implement the SRTF policy for concurrent kernels that uses this predictor and evaluate it on two-program workloads from ERCBench. SRTF improves STP by 1.18x and ANTT by 2.25x over FIFO. Compared to MPMax, a state-of-the-art resource allocation policy for concurrent kernels, SRTF improves STP by 1.16x and ANTT by 1.3x. To improve fairness, we also propose SRTF/Adaptive which controls resource usage of concurrently executing kernels to maximize fairness. SRTF/Adaptive improves STP by 1.12x, ANTT by 2.23x and Fairness by 2.95x compared to FIFO. Overall, our implementation of SRTF achieves STP to within 12.64% of Shortest Job First (SJF, an oracle optimal scheduling policy), bridging 49% of the gap between FIFO and SJF.