Towards Energy Proportional Computing for Enterprise

Question: Discuss about the Towards Energy Proportional Computing for Enterprise. Answer: Towards Energy-Proportional Computing For Enterprise-Class Server Workloads In huge data centers, there are thousands of nodes for computing purpose that also became a common place for enterprise computing processes. The growth of power consumptions of these data centers is occurring at an unprecedented rate. In addition to this, the servers are unable to provide execution, which is energy efficient under every level of utilization, thus do not exhibit an energy proportionality. This also reduces the overall energy level efficiency of these data centers. Hence, effective strategies are needed to improve energy efficiency and for controlling power consumption in data center servers. This report is based on investigating the possibility of achieving the energy proportionality, by using interface of Intel's RAPL (Running Average Power Limit), for server workload of enterprise-class. This report is about analyzing RAPL Interface's impact on per-transaction, performance, energy efficiency and power consumption, under the various level of loads for benchmarking. Intels RAPL Interface The power consumption by typical servers is 35 to 45 percentage of the peak power at idle state. Hence, it is challenging to achieve an energy proportionality for server workload of enterprise-class. In the interface provided by Intel's RAPL, DVFS, which refers to the dynamic scaling of voltage and frequency, is used for achieving improved energy efficiency, since it has the potential to give a cubic amount of energy savings. It is also the part that is responsible for energy proportionality in the system. The other subsystem is uncore which consumes a constant amount of power level, without depending upon system utilization. It has on-chip meters for energy measurement that deals with processor package, DRAM and core subsystem, also enables tracking down the power consumption at a system level basis and time resolution of 1 ms, which was earlier not possible. It also provides control over consumption of power through subsystems power limiting interfaces ("Intel Usage-to-Platform Re quirements Process", 2007). Contributions :Power Management via SPECpower Analysis For server workloads in enterprise-class, RAPL Interface provides a server-side processing application for Java transaction and it is based on the benchmark of SPECjbb2005. It mainly focuses on the CPU, memory hierarchy, caches and also performs implementation tests on JVM, JIT, threads and garbage collection. For benchmark process, it requires SUT and CCS, along with Ethernet that is used for communicating between systems (Stannack, 1996). Performance Characteristics of RAPL Interface Power Limiting - It limits the average power over the sliding window and does not limit instantaneous power. Its advantage is no performance degradation of workload if average performance is within power limits specified. Energy Metering- It reports the information about the actual usage of energy in the system. Performance Status- It reports the impact of performance due to the power limit. There is an MSR Interface that reports the overall time for each domain that functioned below P-state requested by the operating system. It will determine the power limiting effects on particular workloads (He, Jenkins, Wu, 2016). Power Information- It provides a range of values for control attributes that are associated with limiting of power. Each range of values specifies information for a particular domain of RAPL that corresponds to its power consumption limitation. It also includes information such as minimum power and maximum power requirement and maximum time of window ("The power of information", 2016). Power Profiles at Various SPECpower Load Levels SUT and CCS are interconnected via Ethernet gigabit network. The test beds consume an average power of 117 W at the idle condition and 314 W at the load level of 100% of SPECpower. At peak power, the system consumes 37.2%, at idle mode. Impact on Response Characteristics Usually, load levels of 50%, 60% and 70% under SPECpower limits are run by manual configuration. It does not allow to enable operation of energy proportionality at full system limit. But it's possible to achieve power savings moderately at a full system, without degradation of power. Also, with load level decrease, power saving is increased (Ryckbosch, Polfliet, Eeckhout, 2011). Limitation RAPL Interface doesnt evaluate power impact due to memory footprint increase in the benchmark of SPECpower. It doesn't explain the use of limiting memory power in operation of energy proportionality. A multi-dimensional formula for achieving best time window and power limit possible is still required. It doesnt create a system in runtime, using optimization framework and power models. Conclusion The energy and power management remains an issue with the data centers. But efficient management of power in workloads that belonged to the enterprise-class server has great potential to reduce costs that are energy oriented. To improve the energy efficiency of the data centers, there is a need for energy proportionality, which can be achieved by RAPL Interfaces for SPECpower benchmarks. Both the contributions and limitations of Intel's RAPL Interface has been included in this summary report for brief understanding. References He, Y., Jenkins, N., Wu, J. (2016). Smart Metering for Outage Management of Electric Power Distribution Networks.Energy Procedia,103, 159-164. https://dx.doi.org/10.1016/j.egypro.2016.11.266 Intel Usage-to-Platform Requirements Process. (2007).Intel Technology Journal,11(01). https://dx.doi.org/10.1535/itj.1101.03 Ryckbosch, F., Polfliet, S., Eeckhout, L. (2011). Trends in Server Energy Proportionality.Computer,44(9), 69-72. https://dx.doi.org/10.1109/mc.2011.130 Stannack, P. (1996). Purchasing power and supply chain management powertwo different paradigms?a response to Ramsay's Purchasing power (1995).European Journal Of Purchasing Supply Management,2(1), 47-56. https://dx.doi.org/10.1016/0969-7012(95)00021-6 The power of information. (2016).IEEE Transactions On Power Delivery,31(4), 1780-1780. https://dx.doi.org/10.1109/tpwrd.2016.2593671