HPC Clouds and the Energy-Performance Tradeoff
Public cloud platforms have become popular as a means of accessing powerful computing resources without having to make large capital investments.
While cloud computing is not a good fit for all HPC workloads, the lower barrier to entry has had a democratizing effect for some HPC users. Last month, IDC revealed that the percentage of sites using cloud computing to process HPC workloads rose from 13.8 percent in 2011 to 23.5 percent in 2013, with public and private cloud use about equally represented among the 2013 sites.
While some in the HPC community dismissed the suitability of the “cloud” for HPC workloads early on, the grid/cloud/virtualization space has long enjoyed an active research base, a trend that continues to this day. One recent paper addressing this topic from a green angle comes from trio of computer scientists based out of Ho Chi Minh City University of Technology in Vietnam. Their work on the energy efficient allocation of virtual machines in a high performance computing cloud was published this month in the Journal of Science and Technology, Vietnamese Academy of Science and Technology.
While some HPC clouds employ bare metal servers, this study is concerned with the more common type of cloud platform, which uses virtualization technology to provision computational resources in the form of virtual machines (VMs). For cloud datacenters, energy consumption is very often the number one cost center, thus cloud operators are highly motivated to rein in energy use. One way to do this is by deploying energy-efficient management techniques.
These techniques are not without challenges, however. For example, in order to realize an energy efficient resource allocation for virtual machines in an HPC cloud, there is a tradeoff between minimizing the energy consumption of physical machines and satisfying quality of service (e.g., performance or resource availability). Cloud providers can maximize their profit by reducing the power cost by operating the smallest number of physical servers. But, pulling the equation in the opposite direction are cloud customers, who desire the highest performance for their applications.
The situation is further challenged by HPC applications. Resource requirements are application-dependent, but as HPC workloads are mostly CPU-intensive, they are unsuitable for some energy management techniques, such as dynamic consolidation and migration.
In this paper, the researchers propose new VM allocation heuristics that use a metric of performance-per-watt to select the most energy-efficient physical machine for each virtual machine. Their energy-aware schedule algorithm was inspired by the Green500 list’s idea of using a metric (performance-per-watt) to rank energy efficiency. They ask the question: “How can we use a similar metric (e.g. TotalMIPS/Watt) as a criterion for selecting a host on assignment of a new VM and is total energy consumption of the whole system minimum?”
Their technique is called Energy-aware and Performance-per-watt oriented Best-fit (EPOBF). The study compares EPOBF (version 1 and 2) to state-of-the-art heuristics (called PABFD and VBP Greedy) on heterogeneous physical machines, where each machine has a multicore CPU.
The authors selected the most recent version (version 3.0) of CloudSim to model and simulate their HPC cloud and the VM allocation heuristics. Their simulated cloud datacenter has 5,000 heterogeneous PMs and a simulated workload with 29,624 cloudlets, each of which can model an HPC task.
The physical machines break down into one-third HP ProLiant ML110 G5 machines; one-third IBM x3250 machines, and one-third Dell PowerEdge R620 machines. The researchers assume that power consumption of a PM has a linear relationship to CPU utilization.
Energy consumption of the different VM allocation heuristics:
In the setup used for this study, the scheduler does not have access to global information about user jobs and user applications in the future. Users request short-term resources at fixed start times and non-interrupted durations.
The authors conclude that it is possible for the HPC cloud’s scheduler to use the metric of performance-per-watt to allocate VMs to hosts for improved energy-efficiency. The experimental simulations show that the EPOBF heuristics can reduce total energy consumption by 35 percent on average in comparison to the PABFD and VBP Greedy allocation heuristics.
The next step for these computer scientists is to evaluate the performance of EPOBF heuristics on different types of system models and workloads. They also plan to explore the impact of memory in energy models and to develop an accurate power model for multicore PMs.