Fast DRL-based scheduler configuration tuning for reducing tail latency in edge-cloud jobs

Journal of Cloud Computing

Advances, Systems and Applications

Table 6 Percentages of reduced tail latency of AI workloads driven by Alibaba trace 2020 under dynamic resources

Scenario	Job scheduling scenarios			TDP/TRP		DRF _BRA	DRF _LRP	DRF _MRP	GANG _BRA	GANG _DRF _BINPACK	GANG _DRF _BRA	GANG _DRF _LRP	GANG _DRF _MRP	GANG _LRP	GANG _MRP	SLA _BRA	SLA _LRP	SLA _MRP
27	AI workloads	Daytime	Original resources	Discard rate	80%	26.77	25.77	26.41	25.03	25.13	29.61	19.49	33.00	30.27	25.77	29.69	24.18	17.83
28			50% decrease	Discard rate	80%	21.38	24.41	21.29	21.29	21.38	21.38	24.41	21.38	24.41	21.38	21.38	24.59	21.20
29			25% decrease	Discard rate	80%	22.75	24.03	15.64	22.54	15.25	22.64	24.14	15.64	24.03	15.64	22.54	24.03	15.51
30			25% increase	Discard rate	80%	19.28	12.91	31.80	19.28	31.92	19.28	12.72	31.68	12.91	31.92	19.28	12.91	31.80
31			50% increase	Discard rate	80%	20.03	18.75	37.43	20.39	37.43	20.03	18.57	37.32	18.75	37.32	20.39	18.57	37.21
32		Night	Original resources	Discard rate	80%	19.04	17.15	16.95	16.18	16.18	20.96	15.93	14.54	17.15	20.10	20.48	15.22	19.12
33			50% decrease	Discard rate	80%	19.25	22.36	19.15	19.15	19.25	19.25	22.36	19.25	22.36	19.25	19.25	22.53	19.05
34			25% decrease	Discard rate	80%	13.78	15.21	5.85	13.54	5.41	13.66	15.33	5.85	15.21	5.85	13.54	15.21	5.70
35			25% increase	Discard rate	80%	13.93	7.15	27.29	13.93	27.42	13.93	6.94	27.16	7.15	27.42	13.93	7.15	27.29
36			50% increase	Discard rate	80%	16.17	14.84	34.41	16.55	34.41	16.17	14.64	34.29	14.84	34.29	16.55	14.64	34.18