Adaptive Memory Management

1

• Adaptive Memory Management
• in
• HPC Applications
• Paul Slavin
• 9 December 2004

2
The NUMA Environment
3
Scalability Sacrifices Uniformity
4
Implications for Performance

• Additional 'layer' of the memory hierarchy.
• A remote access incurs a greater time cost than a
  local access.
• Memory accesses frequently constitute the largest
  execution overhead.
• Proximity to data becomes a determinant of
  performance.
• Poor locality severely degrades performance.

5
Existing Techniques

• Explicit placement in code. - Programmer
  Placement - Compiler directives, e.g. OpenMP
  (Ext.)
• Compile-time automatic optimisation.
• OS intervention at runtime.

6
Page Migration

• Exploits virtual memory system.
• Reordering of arbitrary V -gt P mappings.
• Performed in response to program behaviour.
• Implemented by redefining TLB ...
• ... or rewriting page table.
• Interaction of hardware with OS or userspace.

7
Migration Illustrated (i)
8
Migration Illustrated (ii)
9
Research Community

• Stanford DASH project.
• UPC NANOS environment.
• Proprietary techniques SGI, AMD, Unisys.
• No clear consensus on page migration per se.

10
Research Avenues

• What are the attributes of successful dynamic
  memory techniques?
• What existing notations can best represent this
  problem?
• - Reconfigurable task systems - Graph theory
• How can these attributes be abstracted from the
  specifics of implementation environments?

11
Research Tools

• Hardware-based profiling - Hardware
  counters - SpeedShop prof
• Valgrind (x86 only)
• Code-based profiling - Timing by system
  call - Memory permissions Unix Signals

12
Implementation

• Improve page migration - Marginal
  gains - Difficult to implement
• Circumvent OS allocation - Speed
  penalty? - Greatest flexibility
• Code analysis - Compiler modification!? - Req
  uires predictable execution?

13
Architectural Awareness

• Similar problems have proved vulnerable to
  sampling of historic performance data (E.g.
  FGDLS)
• Can performance data be used by a program to
  create a representation of its execution
  environment?

14
Bibliography (i)

• 1 R. P. LaRowe Jr., J. T. Wilkes, and C. S.
  Ellis. Exploiting operating system support for
  dynamic page placement on a NUMA shared memory
  multiprocessor. In Proceedings of the 3rd ACM
  SIGPLAN Symposium on Principles Practice of
  Parallel Programming, volume 26, pages 122-132,
  Williamsburg, VA, April 1991.
• 2 Yair Bartal, Moses Charikar, and Piotr
  Indyk. On page migration and other relaxed task
  systems. Theoretical Computer Science,
  268(1)43-66, 2001.
• 3 Jeffery Westbrook. Randomized algorithms for
  multiprocessor page migration. SIAM J. Comput.,
  23(5)951-966, 1994.
• 4 E. T. Roush. Fast dynamic process migration.
  In Proceedings of the 16th International
  Conference on Distributed Computing Systems
  (ICDCS '96), page 637. IEEE Computer Society,
  1996.
• 5 Dongming Jiang and Jaswinder Pal Singh. A
  methodology and an evaluation of the sgi
  origin2000. In Proceedings of the 1998 ACM
  SIGMETRICS joint international conference on
  Measurement and modeling of computer systems,
  pages 171-181. ACM Press, 1998.
• 6 Steve Carr, Kathryn S. McKinley, and
  Chau-Wen Tseng. Compiler optimizations for
  improving data locality. In Proceedings of the
  sixth international conference on Architectural
  support for programming languages and operating
  systems, pages 252-262. ACM Press, 1994.

15
Bibliography (ii)

• 7 Timothy Sherwood, Brad Calder, and Joel
  Emer. Reducing cache misses using hardware and
  software page placement. In Proceedings of the
  13th international conference on Supercomputing,
  pages 155-164. ACM Press, 1999.
• 8 Brad Calder, Chandra Krintz, Simmi John, and
  Todd Austin. Cache-conscious data placement. In
  Proceedings of the eighth international
  conference on Architectural support for
  programming languages and operating systems,
  pages 139-149. ACM Press, 1998.
• 9 Alvin R. Lebeck, Xiaobo Fan, Heng Zeng, and
  Carla Ellis. Power aware page allocation. In
  Proceedings of the ninth international conference
  on Architectural support for programming
  languages and operating systems, pages 105-116.
  ACM Press, 2000.
• 10 Satyendra Bahadur, Viswanathan
  Kalyanakrishnan, and James Westall. An empirical
  study of the effects of careful page placement in
  linux. In Proceedings of the 36th annual
  Southeast regional conference, pages 241-250. ACM
  Press, 1998.
• 11 Jr. Richard P. LaRowe, Mark A. Holliday, and
  Carla Schlatter Ellis. An analysis of dynamic
  page placement on a numa multiprocessor. In
  Proceedings of the 1992 ACM SIGMETRICS joint
  international conference on Measurement and
  modeling of computer systems, pages 23-34. ACM
  Press, 1992.
• 12 Jr. R. P. LaRowe, C. S. Ellis, and M. A.
  Holliday. Evaluation of numa memory management
  through modeling and measurements. IEEE Trans.
  Parallel Distrib. Syst., 3(6)686-701, 1992.