GreenMap with Dynamic Scaling

As is shown by the above result, GreenMap suffers a degradation in response time when the load is below 0.6, and as load grows from low to medium, the degradation decreases. In fact, the lower the load, the more delay Green-Map will apply to tasks in order to emulate a higher load, at which point

there is an abundance of outstanding tasks, hence facilitating assignment in synchronization. We also observe that higher loads can be more efficiently achieved by turning off a fraction of stacks in a large cluster with multiple series-stacks. The preference of higher load by GreenMap provides us with inspirations for further improvement: When the cluster is running at a low load, one can consolidate services to only a fraction of servers and have them running at a higher load, while on the other hand, saving energy from other idle servers by shutting them down.

For instance, assume 10 series-stacks are running at 0.4 load. From Fig-ure 5.2, the total power consumption in each series-stack of 4 servers is 192.2 W (= V P4

i=1I_i) at 0.4 load, and the conversion loss in each series-stack is 0.29 W with GreenMap. With dynamic scaling, we can turn off 3 series-stacks, resulting in 0.57 load for each remaining series-stack. The corre-sponding power consumption in each series-stack is now 215.4 W, and the conversion loss is 0.33 W with GreenMap. Hence, with GreenMap but not dynamic scaling,

total power = (192.2 + 0.29) × 10 = 1924.9W, whereas with GreenMap and dynamic scaling,

total power = (215.4 + 0.33) × 6 = 1294.4W,

which is a 32.8% reduction. The reduction in total energy consumption is similar as the servers are mostly idle at 0.4 and 0.57 load, and the trace takes a similar amount of time to finish. The average job response time with GreenMap will increase by only 15% as the load increases from 0.4 to 0.57, yielding a favorable tradeoff between power efficiency and response time.

CHAPTER 6 CONCLUSION

We explored the feasibility of series-connected stacks in data centers by im-plementing GreenMap, a modified MapReduce framework that assigns tasks in synchronization. We found that with task synchronization, the conver-sion loss in data centers can potentially be reduced by two orders of magni-tude, which is equivalent to about 15% of total energy consumption. Green-Map also achieves 70.1% - 75.4% reduction in conversion loss compared to Hadoop’s FIFO scheduler under serial stack structure. Based on the observa-tion that the average response time of GreenMap suffers a degradaobserva-tion at low load, we further proposed a modification of GreenMap with dynamic scaling to achieve a favorable tradeoff between response time and power efficiency.

Future work includes implementing GreenMap with multiple series-stacks and heterogeneous jobs, and evaluating the system on actual series-connected stacks.

REFERENCES

[1] The Climate Group, “Smart 2020: Enabling the low car-bon economy in the digital age,” 2009. [Online]. Avail-able: https://www.theclimategroup.org/sites/default/files/archive/

files/Smart2020Report.pdf

[2] A. Greenberg, J. Hamilton, D. Maltz, and P. Patel, “The cost of a cloud:

Research problems in data center networks,” ACM Sigcomm Computer Communication Review (CCR), 2009.

[3] J. Whitney and P. Delforge, “Data center efficiency assessment,” Natural Resources Defense Council, Tech. Rep., Aug. 2014.

[4] R. E. Brown, R. Brown, E. Masanet, B. Nordman, B. Tschudi, A. She-habi, J. Stanley, J. Koomey, D. Sartor, P. Chan et al., “Report to congress on server and data center energy efficiency: Public law 109-431,” Ernest Orlando Lawrence Berkeley National Laboratory, Berkeley, CA (US), Tech. Rep., 2007.

[5] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a warehouse-sized computer,” in ISCA’07, 2007.

[6] U. Hoelzle and L. A. Barroso, The Datacenter as a Computer: An In-troduction to the Design of Warehouse-Scale Machines, 1st ed. Morgan and Claypool Publishers, 2009.

[7] A. Krioukov, P. Mohan, S. Alspaugh, L. Keys, D. Culler, and R. Katz,

“Napsac: Design and implementation of a power-proportional web clus-ter,” in Proc. of Sigcomm workshop on Green Networking, 2010.

[8] A. Gandhi, M. Harchol-Balter, R. Raghunathan, and M. Kozuch, “Dis-tributed, robust autoscaling policies for power management in compute intensive server farms,” in Open Cirrus Summit 2011, 2011.

[9] M. Lin, A. Wierman, L. L. Andrew, and E. Thereska, “Dynamic right-sizing for power-proportional data centers,” Proc. IEEE INFOCOM, Shanghai, China, pp. 10–15, 2011.

[10] L. L. H. Andrew, M. Lin, and A. Wierman, “Optimality, fairness, and robustness in speed scaling designs,” in Proc. of Sigmetrics, 2010.

[11] M. Ton, B. Fortenbery, and W. Tschudi, “Dc power for improved data center efficiency,” Lawrence Berkeley National Laboratory, Tech. Rep., 2008.

[12] T. Aldridge, A. Pratt, P. Kumar, D. Dupy, and A. G., “Evaluating 400v direct-current for data centers – a case study comparing 400 vdc with 480-208 vac power distribution for energy efficiency and other benefits.”

Intel Labs, Tech. Rep., Dec. 2012.

[13] E. Candan, P. S. Shenoy, and R. C. Pilawa-Podgurski, “A series-stacked power delivery architecture with isolated differential power conversion for data centers,” in Telecommunications Energy Conference (INT-ELEC), 2014 IEEE 36th International. IEEE, 2014, pp. 1–8.

[14] J. Dean and S. Ghemawat, “MapReduce: Simplified data processing on large clusters,” Commun. ACM, vol. 51, no. 1, pp. 107–113, Jan. 2008.

[Online]. Available: http://doi.acm.org/10.1145/1327452.1327492 [15] Y. Chen, A. Ganapathi, R. Griffith, and R. Katz, “The case for

eval-uating MapReduce performance using workload suites,” in Proc. IEEE Int’l Symp. Modeling, Analysis & Simulation of Comput. Telecomm.

Syst. (MASCOTS), 2011.

[16] S. A. Jyothi, C. Curino, I. Menache, S. M. Narayanamurthy, A. Tu-manov, J. Yaniv, R. Mavlyutov, ´I. Goiri, S. Krishnan, J. Kulkarni et al.,

“Morpheus: Towards automated slos for enterprise clusters.” in OSDI, 2016, pp. 117–134.

[17] C. L. Abad, N. Roberts, K. Lee, Y. Lu, and R. H. Campbell, “A storage-centric analysis of MapReduce workloads: File popularity, temporal lo-cality and arrival patterns,” in IEEE International Symposium on Work-load Characterization (IISWC), 2012.