Dynamic/Adaptive Systems

Cross-cut Group, NSF Workshop on Directions for Systems Research

Writeup: Andrew A. Chien, achien@cs.uiuc.edu

August 1, 1997

Participants:

Andrew A. Chien, University of Illinois at Urbana-Champaign

John Chapin, Massachusetts Institute of Technology

Peter Reiher, University of California at Los Angeles

Frank Anger, National Science Foundation

Bobby Blumofe, University of Texas at Austin

Charter:

The group discussed opportunities for dynamic adaptation to provide more predictable, more robust, higher performance, more composable systems. The focus of the group was on multiprogrammed systems and delivering system performance to real applications. Thus, encompassing issues that involved applications, operating systems, and hardware architecture.

Key conclusion:

Technology trends in hardware and software continue to reduce the hardware and runtime cost of building dynamic, adaptive systems. Concurrently, the increasing dynamic performance range across processors, networks, and storage hierarchies are dramatically increasing the potential benefits of customization. Therefore, dynamic/adaptive systems, that is systems that include significant capability for customization beyond programming in the central processing unit (e.g. in the memory system, I/O system, networking system), are a promising direction for systems research. We believe this area has the potential to deliver quantum leaps in capabilities. Dynamic/adaptive systems can include the following characteristics:

programmable engines in I/O devices
programmable engines in memory controllers
programmable engines in cache controllers
programmable engines in network devices
configurable data paths / logic in any part of the system
configurable queueing in scheduling for interconnect, devices, memory
performance instrumentation and feedback for subsystem utilization
interfaces for information flow amongst application, operating system, and hardware
infrastructure for processing performance data, mechanisms for implementing coordinated resource management
policies for coordinated resource management

Motivating applications/technologies:

rapidly decreasing cost of local computation (gates or transistors)
rapidly increasing cost of communication, relative to local computation
latency-sensitive applications (best effort not good enough)
increasing performance fragility of systems (high Q systems)
increasingly dynamic/unpredictable workloads (software mobility)
excess resources (enable offline analysis and tuning)

Examples where order of magnitude performance benefits are available today:

(these and other opportunities continue to increase)

adaptation of disk layout, based on access pattern
adaptation of virtual memory layout, based on cache miss data
adaptation of resource scheduling (processor, memory, I/O, network), to reduce end to end latency
adaptation of data compression boundaries, based on where data goes into cache
adaptation of memory hierarchy data units, based on data usage, eliminate false sharing
adaptation of application resource usage, based on projection of resource availability
adaptation of application resource acquisition requests, prediction based on past behavior

Examples of Information Flows to support this Adaptation/Configuration:

	Information Flow	Adaptation to Improve System
Application -> OS	Large scale data access patterns	Data organization/cache/prefetching Memory/resource alloc/reservation
	Performance requirements for subsystems	Subsystem scheduling
	Security requirements (as a performance tradeoff)	Selection of encryption boundaries, Transfer paths in the system
Application -> Hardware	Cache performance, Cache protocol control	Optimized data movement
	Dynamic data structures	Custom prefetching and cache block size
	Application data representation	Application specific compression in memory/disk controllers
	Application encryption keys	Application specific encryption in memory/disk controllers
Hardware ->{OS,Application}	Application-specific memory hierarchy behavior	Data Reorganization/Page Mapping
	Queueing and utilization for subsystems	Application-level load management Operating system load management Resource management and scheduling
	Customization specific data -> application	Application-level adaptation for resource and load management, data structures, algorithm, etc.
Network -> {OS, Application}	Load/latency/schedule information (cost or available performance)	WAN resource management, Local application adaptation (e.g. prefetching)

Solutions are not just interfaces, but rather interfaces, information formats, analysis, mechanisms for adaptation, and policies for adaptation. All of these components must be melded together into systems that enable systematic solutions. Information for adaptation can be collected and passed across interfaces explicitly, or it can be derived implicitly (for example by observing previous behavior of the program).

To get research started in this area, the best first steps would be to:

Define initial interfaces for information flow across systems
Define interfaces and mechanisms for dynamic extension, enabling experimentation
Produce a reference compiler, OS, HW systems that implement these interfaces and enable experimentation