Measuring and Exploiting Guardbands of Server-Grade ARMv8 CPU Cores and DRAMs

Autor: Tovletoglou, Konstantinos, Mukhanov, Lev, Karakonstantis, Georgios, Chatzidimitriou, Athanasios, Papadimitriou, George, Kaliorakis, Manolis, Gizopoulos, Dimitris, Hadjilambrou, Zacharias, Sazeides, Yiannakis, Lampropulos, Alejandro, Das, Shidhartha, Vo, Phong
Přispěvatelé: Tovletoglou, Konstantinos [0000-0002-1513-3143]
Jazyk: angličtina
Rok vydání: 2018
Předmět:
Zdroj: Tovletoglou, K, Mukhanov, L, Karakonstantis, G, Chatzidimitriou, A, Papadimitriou, G, Kaliorakis, M, Gizopoulos, D, Hadjilambrou, Z, Sazeides, Y, Lampropoulos, A, Das, S & Vo, P 2018, Measuring and Exploiting Guardbands of Server-Grade ARMv8 CPU Cores and DRAMs . in 2018 48th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN): Proceedings . pp. 6-9, IEEE/IFIP International Conference on Dependable Systems and Networks, Luxembourg, Luxembourg, 25/06/2018 . https://doi.org/10.1109/DSN-W.2018.00013
DSN Workshops
2018 48th Annual IEEE/IFIP International Conference on Dependable Systems and Networks Workshops (DSN-W)
DOI: 10.1109/DSN-W.2018.00013
Popis: In this paper, we present the results of our comprehensive measurement study of the timing and voltage guardbands in memories and cores of a commodity ARMv8 based micro-server. Using various synthetic micro-benchmarks, we reveal how the adopted voltage margins vary among the 8 cores of the CPU chip, and among 3 different sigma chips and we show how prone they are to worst-case voltage noise. In addition, we characterize the variation of 'weak' DRAM cells in terms of their retention time across 72 DRAM chips and evaluate the error mitigation efficacy of the available error-correcting codes in case of operation under aggressively relaxed refresh periods. Finally, we show the overall energy savings that could be achieved by shaving the adopted guardbands in the cores and memories using various applications. Our characterization results show the potential to obtain up-to 38.8% energy savings in cores and up-to 27.3% within DRAMs. 6 9
Databáze: OpenAIRE