JavaTile: CMP-simulation with a twist

1
JavaTile CMP-simulation with a twist

• Dan Greenfield
• Computer Architecture Group

Internal Presentation, 16th February 2007
2
Aim of Talk

• Introduce JavaTile
• Show benefits and problems of approach
• Spark interest in collaboration
• Invite expertise from multiple areas to solve CMP
  problems

3
Quick Background Exciting Times!
Cisco 188-core (50 BIPS) 2
Intel 80-core (1 TFLOPS) 1
4
Parts of a CMP

• Q How well do each of the components run?
• Q How well does the network run?

From Pestata et al 2004 3
5
Parts of a CMP continued

• Real Q How well do Applications run?

6
Motivations

• Need more realistic NoC traffic
• Current methods synthetic, limited applications,
  low PE count, course-grain, OO Superscalar
  internals
• How is the network used?
• What is needed in NoC for future CMP?
• Want System-level view of performance, power and
  fault-tolerance
• Most current metrics concern the NoC and 'guess'
  what this means for the system-level
• Want to explore solutions at all levels

7
Some Existing CMP Approaches

• SimpleScalar-based CMP simulator
• Hydra 4 MIPS-core CMP simulator
• CMP-SIM (extension of SimpleScalar)
• SESC Superscalar (1.5MIPS on 3GHz P4)
• GEMS (commercial SIMICS-based)
• ML-RSIM (Sparc RSIM-based)

8
Java Virtual Machine

• Platform with standard library
• Virtual Processor executing Java instruction set
  'bytecode'
• Compilable to native platform

9
Java Advantages

• A widely deployed standard platform
• Its 'machine code' is itself Object Oriented with
  type information
• Amenable to static code analysis
• Tools to run efficiently, or compile to native
  executable

10
JavaTile Processing Element
11
JavaTile System
12
Bytecode Instrumentation

• Hook into all instructions that may cause NoC
  traffic

Fibonacci2() Code 0 bipush 0 2
bipush -33 4 invokestatic 23
//Method monitor/Monitor.methodStart(II)V 7
sipush -29729 10 sipush 0 13
invokestatic 26 //Method monitor/Monitor.jum
pMarker(II)V 16 aload_0 17 sipush
1 20 invokestatic 30 //Method
monitor/Monitor.syncCycleCount(I)V 23
invokespecial 32 //Method java/lang/Object."lt
initgt"()V 26 sipush -29729 29
sipush 4 32 invokestatic 35
//Method monitor/Monitor.postMethodCall(II)V
35 return
13
Current Flow
14
Problems

• Garbage Collection
• Local memory vs global memory allocation
• Passing by pointers (ownership)
• Push versus Pull
• No Inlining
• Auto-Parallelization
• Debugging

15
Auto-Parallelization

• Software Pipelining
• e.g. MIT RAW Compiler 4
• e.g. Princeton DSWP (Decoupled SWP) 5
• Thread-Level Speculation
• Loop-level (e.g. Stanford Jrpm) 6
• Method-level (e.g. SableSpMT) 7
• Affine Partitioning
• e.g. Incorporated in Stanford SUIF 8

16
References

• 1 Intel Polaris, from IDF 2006 slides, photo at
  http//www.tomshardware.com
• 2 W. Eatherton, The Push of Network Processing
  to the Top of the Pyramid, Keynote Slides at
  http//www.cesr.ncsu.edu/ancs/slides/eatherton/Key
  note.pdf
• 3 Pestata et al, Cost-Performance Trade-Offs in
  Networks on Chip A Simulation-Based Approach,
  DATE 2004
• 4 Waingold et al, Baring it All to Software
  Raw Machines, Computer Vol 30, 9, 1997
• 5 Ottoni et al, Automatic Thread Extraction
  with Decoupled Software Pipelining. MICRO 2005
• 6 Chen et al, The Jrpm System for Dynamically
  Parallelizing Sequential Java Programs, IEEE
  Micro Vol 23, No 6, Nov/Dec 2003
• 7 Pickett et al, SableSpMT a software
  framework for analysing speculative
  multithreading in Java, PASTE Workshop 2006
• 8 Lim et al, An affine partitioning algorithm
  to maximize parallelism and minimize
  communications, ACM SIGARCH 1999