Adaptive performance prediction for integrated GPUs

Gupta, Ujjwal; Campbell, Joseph; Ogras, Ümit Y.; Ayoub, Raid; Kishinevsky, Michael; Paterna, Francesco; Gümüşsoy, Suat

doi:10.1145/2966986.2966997

Cited by 23 publications

(28 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Predictive models help to tune performance for varying workload. Gupta et al (2016) predicts GPU performance online as DPM algorithm tunes GPU frequency. Proactive techniques are enabled by predictive models and achieve better quality of control than reactive methods.…”

Section: Observing and Predictingmentioning

confidence: 99%

On-Chip Dynamic Resource Management

Miele

Kanduri²,

Moazzemi³

et al. 2019

FNT in Electronic Design Automation

View full text Add to dashboard Cite

The need for dynamic resource management has shadowed the exponential growth of on-chip transistor capacity, and the challenge is accentuated by the heterogeneity of resources, and the bewildering variety of constraints and requirements of applications, platforms and users. The field has started with a few research papers in the early 1990s but has grown today to over hundred yearly publications, leading to an accumulated body of literature presumable far above 1000 papers. We focus on the dynamic (run-time) management of onchip resources and mostly ignore design-time techniques and off-chip resources of larger electronic systems. Moreover, we do not attempt a complete review of all published work on the topic. Rather, this survey provides a structured review and discussion of the state of the art and is divided along the primary objectives of resource management techniques: performance, power, reliability and quality of service, each

show abstract

Section: Observing and Predictingmentioning

confidence: 99%

On-Chip Dynamic Resource Management

Miele

Kanduri²,

Moazzemi³

et al. 2019

FNT in Electronic Design Automation

View full text Add to dashboard Cite

show abstract

“…Unlike our work, these models are not built for GPUs, and do not use frequency and performance counters as inputs. Our prior performance model for integrated GPUs [11] also employ RLS algorithm and performance counters. However, it requires offline learning to characterize the frequency dependence of the counters used by the RLS algorithm.…”

Section: Related Researchmentioning

confidence: 99%

“…This equation can be used to predict the change in frame time for the new candidate frequency as: ∂ f , which is zero for frequency independent counters. Note that our prior work employed a non-linear offline model to compute [11]. It is possible to learn this model online as well by employing two parallel adaptive algorithms, but that will incur more overhead.…”

Section: B Frame Time Sensitivitymentioning

confidence: 99%

“…The second algorithm employs the same model with online learning using the DCD-RLS algorithm (DCD-RLS). The third algorithm employs two models: (a) the model shown in Equation 6 with online learning using RLS and (b) an offline nonlinear model for derivative of frequency dependent counter with respect to frequency (RLS+Offline) [11].…”

Section: A Experimental Setupmentioning

confidence: 99%

“…Neither of these counters depend on the frequency; they are functions of only the frame complexity. Note that in our prior work [11] we also select four features, but these consist of a single frequency change term and three counters. Two of these counters, Aggregate Core Array Active and Slow Z Test Pixels Failing are frequency independent and one counter Rendering Engine Busy is frequency dependent.…”

Section: B Offline Feature Selection and 2 Regularizationmentioning

confidence: 99%

See 2 more Smart Citations

An Online Learning Methodology for Performance Modeling of Graphics Processors

Gupta

Babu²,

Ayoub³

et al. 2018

IEEE Trans. Comput.

Self Cite

View full text Add to dashboard Cite

Approximately 18% of the 3.2 million smartphone applications rely on integrated graphics processing units (GPUs) to achieve competitive performance. Graphics performance, typically measured in frames per second, is a strong function of the GPU frequency, which in turn has a significant impact on mobile processor power consumption. Consequently, dynamic power management algorithms have to assess the performance sensitivity to the frequency accurately to choose the operating frequency of the GPU effectively. Since the impact of GPU frequency on performance varies rapidly over time, there is a need for online performance models that can adapt to varying workloads. This paper presents a light-weight adaptive runtime performance model that predicts the frame processing time of graphics workloads at runtime without apriori characterization. We employ this model to estimate the frame time sensitivity to the GPU frequency, i.e., the partial derivative of the frame time with respect to the GPU frequency. The proposed model does not rely on any parameter learned offline. Our experiments on the Intel Minnowboard MAX platform running common GPU benchmarks show that the mean absolute percentage error in frame time and frame time sensitivity prediction are 4.2% and 6.7%, respectively.

show abstract