A performance-aware quality of service-driven scheduler for multicore processors

Abstract

In the latest decade, the IT industry shifted from single to multicore processors. Multicore processors require better support from operating systems and runtimes to allow applications to achieve predictable performance and guarantee quality of service (QoS). Finding a proper schedule to yield the specified performance for single and multi-threaded applications can be cumbersome; dealing with multi-programmed workloads may be even worse.We present a performance-aware QoS-driven scheduler for multicore processors, which exploits the availability of runtime application-specific performance measurements to determine a suitable allotment of cores for multi-programmed workloads so as to achieve the desired level of QoS. The proposed scheduler is meant to be implemented in user-mode and harnesses an auto-regressive moving average performance model to put in a relationship performance measurements and resource allocation and is capable of embodying applications' characteristics such as execution phases.

Automated Fine-Grained CPU Provisioning for Virtual Machines

“…AutoPro requires each SLO-bound VM to make periodic performance reports available to its controller, in order to leverage its resource-performance models, as proposed in previous works [Zhang et al 2002;Padala et al 2009;Shen et al 2011;Sironi et al 2012;Bartolini et al 2013a;Hoffmann et al 2013;Sironi et al 2014]. Any performance metric meaningful to the user can be used for these reports and to express SLOs; for instance, a web server can use throughput (e.g., requests/s for a web server) or latency (i.e., response time).…”

“…While these benchmarks were not designed to capture all the characteristics of typical cloud workloads, colocating PARSEC applications does create contention on compute bandwidth, thus stressing the problem that this article addresses. Since PARSEC applications do not natively report performance at runtime, we instrument a subset of the suite 5 to report throughput through our efficient user-space implementation of the Application Heartbeats API [Hoffmann et al 2010;Sironi et al 2012;Bartolini et al 2013a;Sironi et al 2014]. The hypervisor accesses VM performance measurements as in previous work [Padala et al 2009;Shen et al 2011].…”

“…To give an example of how this metric can characterize the performance of PARSEC applications, Figure 3 shows the traces of two VMs with radically different behavior: with a constant resource allocation, swaptions has a stable performance throughout its execution, while x264 shows input-dependent 6 oscillations [Sasaki et al 2012;Bartolini et al 2013a;Sironi et al 2014]. Measuring throughput over an adequate sliding window allows the controllers to promptly catch transitions between execution phases and adjust resource allocations as needed (see Section 4.3).…”

“…Since this article focuses on managing VMs that run compute-intensive multithreaded tasks, we use the cpuset and cpu [Turner et al 2010] subsystems to allocate fractions of the CPU. These two subsystems control CPU partitioning and multiplexing: the cpuset subsystem allows one to pin threads (in our case, vCPUs) in a cgroup to a subset of the available cores, while the cpu subsystem allows one to set a cap on the CPU bandwidth each cgroup can use [Turner et al 2010]. 7 AutoPro can support both CPU partitioning and multiplexing; we evaluated both mechanisms on managing VMs executing multithreaded applications from the PARSEC 2.1 benchmark suite colocated on our hexa-core host node (configuration details are in Section 4).…”

“…We verify that these applications scale well to six threads [Sasaki et al 2012] and configure VMs with six vCPUs. To allocate the host CPU, we either pack vCPUs on a subset of the host cores or cap the host CPU bandwidth [Turner et al 2010]. …”