Canturk Isci

An Analysis of Efficient Multi-Core Global Power Management Policies: Maximizing Performance for a Given Power Budget

Chip-level power and thermal implications will continue to rule as one of the primary design constraints and performance limiters. The gap between average and peak power actually widens with increased levels of core integration. As such, if per-core control of power levels (modes) is possible, a global power manager should be able to dynamically set the modes suitably. This would be done in tune with the workload characteristics, in order to always maintain a chip-level power that is below the specified budget. Furthermore, this should be possible without significant degradation of chip-level throughput performance. We analyze and validate this concept in detail in this paper. We assume a per-core DVFS (dynamic voltage and frequency scaling) knob to be available to such a conceptual global power manager. We evaluate several different policies for global multi-core power management. In this analysis, we consider various different objectives such as prioritization and optimized throughput. Overall, our results show that in the context of a workload comprised of SPEC benchmark threads, our best architected policies can come within 1% of the performance of an ideal oracle, while meeting a given chip-level power budget. Furthermore, we show that these global dynamic management policies perform significantly better than static management, even if static scheduling is given oracular knowledge

Runtime power monitoring in high-end processors: methodology and empirical data

With power dissipation becoming an increasingly vexing problem across many classes of computer systems, measuring power dissipation of real, running systems has become crucial for hardware and software system research and design. Live power measurements are imperative for studies requiring execution times too long for simulation, such as thermal analysis. Furthermore, as processors become more complex and include a host of aggressive dynamic power management techniques, per-component estimates of power dissipation have become both more challenging as well as more important. In this paper we describe our technique for a coordinated measurement approach that combines real total power measurement with performance-counter-based, per-unit power estimation. The resulting tool offers live total power measurements for Intel Pentium 4 processors, and also provides power breakdowns for 22 of the major CPU subunits over minutes of SPEC2000 and desktop workload execution. As an example application, we use the generated component power breakdowns to identify program power phase behaviour. Overall, this paper demonstrates a processor power measurement and estimation methodology and also gives experiences and empirical application results that can provide a basis for future power-aware research.

Efficient resource provisioning in compute clouds via VM multiplexing

Resource provisioning in compute clouds often require an estimate of the capacity needs of Virtual Machines (VMs). The estimated VM size is the basis for allocating resources commensurate with workload demand. In contrast to the traditional practice of estimating the VM sizes individually, we propose a joint-VM sizing approach in which multiple VMs are consolidated and provisioned, based on an estimate of their aggregate capacity needs. This new approach exploits statistical multiplexing among the workload patterns of multiple VMs, i.e., the peaks and valleys in one workload pattern do not necessarily coincide with the others. Thus, the unused resources of a low utilized VM can be directed to the other co-located VMs with high utilization. Compared to individual VM based provisioning, joint-VM sizing and provisioning may lead to much higher resource utilization. This paper presents three design modules to enable the concept in practice. Specifically, a performance constraint describing the capacity need of a VM for achieving a certain level of application performance; an algorithm for estimating the size of jointly provisioning VMs; a VM selection method that seeks to find good VM combinations for being provisioned together. We showcase that the proposed three modules can be seamlessly plugged into existing applications such as resource provisioning, and providing resource guarantees for VMs. The proposed algorithms and applications are evaluated by monitoring data collected from about 16 thousand VMs in commercial data centers. These evaluations reveal more than 45% improvements in terms of the overall resource utilization.

Long-term workload phases: duration predictions and applications to DVFS

Computer systems increasingly rely on adaptive dynamic management of their operations to balance power and performance goals. Such dynamic adjustments rely heavily on the systems ability to observe and predict workload behavior and system responses. The authors characterize the workload behavior of full benchmarks running on server-class systems using hardware performance counters. Based on these characterizations, they developed a set of long-term value, gradient, and duration prediction techniques that can help systems to provision resources.

Identifying program power phase behavior using power vectors

Characterizing program behavior is important for both hardware and software research. Most modern applications exhibit distinctly different behavior throughout their runtimes, which constitute several phases of execution that share a greater amount of resemblance within themselves compared to other regions of execution. These execution phases can occur at very large scales, necessitating prohibitively long simulation times for characterization. Due to the implementation of extensive clock gating and additional power and thermal management techniques in modern processors, these program phases are also reflected in program power behavior, which can be used as an alternative means of program behavior characterization for power-oriented research. In this paper, we present our methodology for identifying phases in program power behavior and determining execution points that correspond to these phases, as well as defining a small set of power signatures representative of overall program power behavior. We define a power similarity metric as an intersection of both magnitude based and ratio-wise similarities in the power dissipation of processor components. We then develop a thresholding algorithm in order to partition the power behavior into similarity groups. We illustrate our methodology with the gzip benchmark for its whole runtime and characterize gzip power behavior with both the selected execution points and defined signature vectors.

Runtime Demand Estimation for effective dynamic resource management

Systems management techniques that allocate resources to running entities, such as processes and virtual machines (VMs), often require estimates of the resources required by each of these resource consumers. For example, many proposed virtual machine placement algorithms attempt to allocate VMs to physical hosts in such a way as to minimize the number of physical hosts that are occupied, while ensuring that each VM receives the CPU required to do its task adequately. The common practice is to assume that the CPU requirement is equal to the current CPU utilization, or to use a prediction of it over an appropriate time horizon. In this paper, we demonstrate that, when multiple VMs or processes co-reside on a physical host, the measured CPU utilization may provide a poor estimate of the actual requirement. We derive a simple, much more accurate alternative estimate of CPU demand, implement it, and demonstrate its superiority experimentally. Furthermore, we demonstrate that using our demand estimation framework in conjunction with dynamic resource allocation in a virtualized environment greatly improves the effectiveness of dynamic placement, resulting in one-shot convergence to optimal placement and significant improvements in the overall performance of the individual VMs.

Improving server utilization using fast virtual machine migration

Live virtual machine (VM) migration is a technique for transferring an active VM from one physical host to another without disrupting the VM. In principle, live VM migration enables dynamic resource requirements to be matched with available physical resources, leading to better performance and reduced energy consumption. However, in practice, the resource consumption and latency of live VM migration reduce these benefits to much less than their potential. We demonstrate how these overheads can be substantially reduced, enabling live VM migration to fulfill its promise. Specifically, we first experimentally study several factors that contribute to the resource consumption and latency of live VM migration, including workload characteristics, the hypervisor and migration configuration, and the available system and network resources. Then, from the insights gained, we propose an alternative remote direct memory access-based migration technique that significantly reduces VM migration overheads. Finally, via simulation and experiments with real system prototypes, we demonstrate that the reduced VM migration overhead results in significant improvements in resource and energy efficiencies, relative to existing migration techniques.

Agile, efficient virtualization power management with low-latency server power states

One of the main driving forces of the growing adoption of virtualization is its dramatic simplification of the provisioning and dynamic management of IT resources. By decoupling running entities from the underlying physical resources, and by providing easy-to-use controls to allocate, deallocate and migrate virtual machines (VMs) across physical boundaries, virtualization opens up new opportunities for improving overall system resource use and power efficiency. While a range of techniques for dynamic, distributed resource management of virtualized systems have been proposed and have seen their widespread adoption in enterprise systems, similar techniques for dynamic power management have seen limited acceptance. The main barrier to dynamic, power-aware virtualization management stems not from the limitations of virtualization, but rather from the underlying physical systems; and in particular, the high latency and energy cost of power state change actions suited for virtualization power management. In this work, we first explore the feasibility of low-latency power states for enterprise server systems and demonstrate, with real prototypes, their quantitative energy-performance trade offs compared to traditional server power states. Then, we demonstrate an end-to-end power-aware virtualization management solution leveraging these states, and evaluate the dramatically-favorable power-performance characteristics achievable with such systems. We present, via both real system implementations and scale-out simulations, that virtualization power management with low-latency server power states can achieve comparable overheads as base distributed resource management in virtualized systems, and thus can benefit from the same level of adoption, while delivering close to energy-proportional power efficiency.

Towards data center self-diagnosis using a mobile robot

We describe an inexpensive robot that serves as a physical autonomic element, capable of navigating, mapping and monitoring data centers with little or no human involvement, even ones that it has never seen before. Through a series of real experiments and simulations, we establish that the robot is sufficiently accurate, efficient and robust to be of practical benefit in real data center environments. We demonstrate how the robots integration with Maximo for Energy Optimization, a commercial data center energy management product, supports autonomic management at the level of the data center as a whole, particularly self-diagnosis of emerging thermal problems.

Robotic mapping and monitoring of data centers

We describe an inexpensive autonomous robot capable of navigating previously unseen data centers and monitoring key metrics such as air temperature1. The robot provides real-time navigation and sensor data to commercial IBM software, thereby enabling real-time generation of the data center layout, a thermal map and other visualizations of energy dynamics. Once it has mapped a data center, the robot can efficiently monitor it for hot spots and other anomalies using intelligent sampling. We demonstrate the robots effectiveness via experimental studies from two production data centers.