100 likes | 227 Vues
This document provides an in-depth characterization of various existing systems within the Department of Computer Sciences at UT Austin. It outlines key applications, components, and policies related to different system architectures, including SEDA, StagedServer, Cilk, and others. Each system's structure, including components like processors and threads, is analyzed along with the context and monitoring mechanisms employed. The report emphasizes policy abstractions and scheduling strategies to enhance performance and resource utilization across platforms.
E N D
Characterization of Existing Systems Puneet Chopra Ramakrishna Kotla Department of Computer Sciences, UT Austin
SEDA Applications • Component SEDA stage • Stage SEDA stage • Path Application/user defined paths within applications • PE Processor • Context Event • Thread Kernel threads in thread pool associated with each stage • Monitor Stage queue length (implicit) • Policy Abstraction • C S : Mandatory • T S : Thread pool + mechanism to adjust pool size based on queue back pressure. • PE T : Time slicing • C T : FIFO/ user defined schedulers/ batching Department of Computer Sciences, UT Austin
StagedServer • Component Stage • Stage Stage • Path Application built over SS library • PE Processor • Context Instance of Closure class • Thread Associated with each PE • Monitor Processor queue length (implicit) • Policy Abstraction At different levels • C S : Mandatory • T S : Wavefront algorithm. • PE S : T S • PE T : Space Splicing • C T : Cohort Scheduling Department of Computer Sciences, UT Austin
Cilk • Component Non replicated node of dag • # of components = 2*(# of cilk fns) + # of syncs. • Stage Spawned thread • Path Runtime dag • PE Processor • Context Arguments passed to spawned threads. • Thread Associated with each PE • Monitor Work queue length (implicit) • Policy Abstraction • C S : Push bottom (fast clone) and Mandatory (slow clone) • T S : Work Stealing. • PE S : T S • PE T : Space Splicing • C T : LIFO Department of Computer Sciences, UT Austin
Hood • Component Class inheriting HoodThread • Stage Spawned thread • Path Application • PE Processor • Context Arguments passed to HoodThreads. • Thread HoodSerMach objects, specified at execution time. • Monitor Work queue length (implicit) • Policy Abstraction • C S : Mandatory • T S : Work Stealing. • PE S : No explicit policy • PE T : Time Splicing • C T : FIFO because only one context per stage Department of Computer Sciences, UT Austin
Coign • Component COM object • Stage Collection of COM objects instantiated on a machine. • Path Application being partitioned • PE Machine • Context COM method invocation parameters • Thread One or many associated with a PE. • Monitor RTE imposed interface. • Policy Abstraction • C S : Mandatory • T S : • PE S : Space partitioning • PE T : • C T : FIFO Department of Computer Sciences, UT Austin
JAWS • Component Class methods • Stage Framework, patterns, class libraries or objects. • Path JAWS server. • PE Processor • Context Request context • Thread Thread • Monitor None to our knowledge. • Policy Abstraction Service configurator pattern. • C S : Mandatory • T S : Multithreading • PE S : • PE T : Time Splicing • C T : FIFO + thread pool/thread per request/single thread Department of Computer Sciences, UT Austin
Click Modular Router • Component Element • Stage Element • Path Packet flow path in router graph • PE Processor (current design for uniprocessor) • Context Packet • Thread One, associated with the PE. • Monitor Queue length, throughput, latency etc. • Policy Abstraction Hierarchical CPU scheduling. • C S : Mandatory • T S : Task queue. • PE S : No explicit policy • PE T : Space partitioning • C T : any scheduler (FIFO etc) Department of Computer Sciences, UT Austin
Scout OS • Component Router • Stage Router • Path Scout Path • PE Processor • Context Message • Thread Thread • Monitor None • Policy Abstraction Thread scheduling on paths. • C S : Mandatory • T S : Multithreading. • PE S : No explicit policy • PE T : Time/Space splicing • C T : EDF and Fixed priority RR Department of Computer Sciences, UT Austin