Real-Time GPU Resource Management with Loadable Kernel Modules

Suzuki, Yuhei; Fujii, Yusuke; Azumi, Takuya; Nishio, Naomichi; Kato, Shigeo

doi:10.1109/tpds.2016.2630697

Cited by 6 publications

(4 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For discrete CPU-GPU systems the point of interference is reduced to data movements between the discrete CPU and GPU memories over the PCIe bus. In this context, several approaches have been proposed [35] [36] [37]. Concerning integrated GPUs, the point of interference is not limited to a single peripheral bus, as the memory hierarchy itself is largely shared between the two devices.…”

Section: Related Workmentioning

confidence: 99%

HePREM: A Predictable Execution Model for GPU-based Heterogeneous SoCs

Forsberg

Benini

Marongiu

2021

IEEE Trans. Comput.

View full text Add to dashboard Cite

The ever-increasing need for computational power in embedded devices has led to the adoption of heterogeneous SoCs combining a general purpose CPU with a data parallel accelerator. These systems rely on a shared main memory (DRAM), which makes them highly susceptible to memory interference. A promising software technique to counter such effects is the Predictable Execution Model (PREM). PREM ensures robustness to interference by separating programs into a sequence of memory and compute phases, and by enforcing a platform-level schedule where only a single processing subsystem is permitted to execute a memory phase at a time. This paper demonstrates for the first time how PREM can be applied to heterogeneous SoCs, based on a synchronization technique for memory isolation between CPU and GPU plus a compiler to transform GPU kernels into PREM-compliant codes. For compute bound GPU workloads sharing the DRAM bandwidth 50/50 with the CPU we guarantee near-zero timing varibility at a performance loss of just 59%, which is one to two orders of magnitude smaller than the worst case we see for unmodified programs under memory interference.

show abstract

Section: Related Workmentioning

confidence: 99%

HePREM: A Predictable Execution Model for GPU-based Heterogeneous SoCs

Forsberg

Benini

Marongiu

2021

IEEE Trans. Comput.

View full text Add to dashboard Cite

show abstract

“…When scanning IP pairs, SRLA only sets some bits of SEA to 1 or sets some DR to 0 and the results are the same regardless of the sequence of execution. Therefore, large amount of IP pairs can be processed simultaneously by updating SEA and DR via multiple threads [37].…”

Section: A Deploy On Gpumentioning

confidence: 99%

A Super Point Detection Algorithm Under Sliding Time Windows Based on Rough and Linear Estimators

Ding

Gong

et al. 2019

IEEE Access

View full text Add to dashboard Cite

Detecting super points from high-speed networks can effectively help to monitor networks, which is a hot topic in network fields. Most existing algorithms are carried out under discrete time windows and the results are in a certain percentage of omission. In this paper, the phenomenon of missed super points detection in discrete time windows is analyzed based on real-world traffic. Then a new algorithm, which detects the super points under sliding time windows, is proposed. Our algorithm uses a lightweight estimator to identify candidate super points and a linear estimator to filter super points. The lightweight estimator is fast, and the linear estimator has high accuracy. Both the lightweight estimator and the linear estimator adopt a data structure, called distance recorder, to support sliding time windows. Moreover, our algorithm is also a parallel algorithm. On the basis of thoroughly discussing the mathematic principles and operation steps of our algorithm, two groups of real-world traffic from a 40-Gb/s high-speed network are applied in the experiments which running on a graphic processing unit (GPU). The experiments are conducted under discrete time windows and sliding time windows separately. The former results show that our algorithm is superior to other existing algorithms in the comprehensive performance, and the latter results indicate that our algorithm can run steadily under sliding time windows.

show abstract

“…Some familiar metrics are IPC (Instructions Per Cycle) and cache_hit_rate. There are several metrics in CUPTI that reflect resource utilization, e.g., dram_utilization, L2_utilization, (11) scan (18) matrixMul (20) mergeSort (14) histogram (15) sortingNetworks (18) vectorAdd (18) transpose (19) Rodinia Pathfinder (10) bfs (16) backprop (8) nn (12) b+ (4) srad (4) hotspot (12) kmeans (20) alu_fu_utilization, or metrics that correspond to the size of data transferred among the resources of a GPU, such as local_load_transactions, and dram_write_transactions.…”

Section: Kernel Metricsmentioning

confidence: 99%

“…Therefore, one of the main targets of the proposed method is the scheduling layer of such systems. There already exist different scheduling methods, such as [9][10][11]. However, they rely on modifications in hardware specifications [12] while our proposed method is applied at software level.…”

Section: Introductionmentioning

confidence: 99%

Metric Selection for GPU Kernel Classification

Shekofteh

Noori

Naghibzadeh

et al. 2018

ACM Trans. Archit. Code Optim.

View full text Add to dashboard Cite

Graphics Processing Units (GPUs) are vastly used for running massively parallel programs. GPU kernels exhibit different behavior at runtime and can usually be classified in a simple form as either "compute-bound" or "memory-bound." Recent GPUs are capable of concurrently running multiple kernels, which raises the question of how to most appropriately schedule kernels to achieve higher performance. In particular, coscheduling of compute-bound and memory-bound kernels seems promising. However, its benefits as well as drawbacks must be determined along with which kernels should be selected for a concurrent execution. Classifying kernels can be performed online by instrumentation based on performance counters. This work conducts a thorough analysis of the metrics collected from various benchmarks from Rodinia and CUDA SDK. The goal is to find the minimum number of effective metrics that enables online classification of kernels with a low overhead. This study employs a wrapper-based feature selection method based on the Fisher feature selection criterion. The results of experiments show that to classify kernels with a high accuracy, only three and five metrics are sufficient on a Kepler and a Pascal GPU, respectively. The proposed method is then utilized for a runtime scheduler. The results show an average speedup of 1.18× and 1.1× compared with a serial and a random scheduler, respectively.

show abstract

Real-Time GPU Resource Management with Loadable Kernel Modules

Cited by 6 publications

References 31 publications

HePREM: A Predictable Execution Model for GPU-based Heterogeneous SoCs

HePREM: A Predictable Execution Model for GPU-based Heterogeneous SoCs

A Super Point Detection Algorithm Under Sliding Time Windows Based on Rough and Linear Estimators

Metric Selection for GPU Kernel Classification

Contact Info

Product

Resources

About