CPI2: CPU performance isolation for shared compute clusters

By Xiao Zhang, Eric Tune, R. Hagmann, Rohit Jnagal, Vrigo Gokhale, John Wilkes. European Conference on Computer Systems, 2013

357 citations recorded by the source corpus.

Subfields: Computer Networks and Communications, Hardware and Architecture, Information Systems

View Paper: https://www.semanticscholar.org/paper/8df87f9f18dc58c2cc88474c2b9e1c30654c60e8

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Networks and Communications

In Hardware and Architecture

In Information Systems

Citations in the corpus

SoftSKU: Optimizing Server Architectures for Microservice Diversity @Scale (2019, background): “Other works reduce coscheduled job interference [109–114] or schedule them in a machine characteristics-aware manner [115–118].”
Large-scale cluster management at Google with Borg (2015, methodology): “The data was gathered from ∼ 12000 randomly selected prod tasks over a week, counting cycles and instructions over a 5 minute interval using the hardware proﬁling infrastructure described in [83], and weighting samples so that every second of CPU time is counted equally.”
Apollo: Scalable and Coordinated Scheduling for Cloud-Scale Computing (2014, neutral): “Existing work leverages Linux containers [13], usage monitoring [27] and/or contention detection [31] to provide performance isolation.”
Heracles: Improving resource efficiency at scale (2015, methodology): “Many previous approaches use IPC instead of latency as the performance metric [79, 51, 60, 80].”
PARTIES: QoS-Aware Resource Partitioning for Multiple Interactive Services (2019, background): “Alternatively, if multiple interactive applications are co-scheduled on a physical host, their load is dialed down considerably, leading to underutilization [61].”
Quasar: resource-efficient and QoS-aware cluster management (2014, background): “Finally, CPI2 [73] throttles low-priority workloads that induce interference to important services.”
Shenango: Achieving High CPU Efficiency for Latency-sensitive Datacenter Workloads (2019, methodology): “For example, Microsoft Bing colocates latency-sensitive and batch jobs on over 90,000 servers [34], and the median machine in a Google compute cluster runs eight applications [76].”
Twine: A Unified Cluster Management System for Shared Infrastructure (2020, background): “Two factors lead to our decision of adopting small machines: 1) our legacy large services were optimized for utilizing entire machines running in private pools, and 2) our stacking technology needed to mature and improve support for performance isolation [42].”
Autopilot: workload autoscaling at Google (2020, background): “Unlike traditional fair-sharing at the OS level, this ensures that a task’s performance is consistent across different executions of the same binary, different machines (as long as the hardware is the same) and different neighbors (tasks co-scheduled on the same machine) [38].”