On a GridBased Interface to a SpecialPurpose Hardware Cluster

1
On a Grid-Based Interface to a Special-Purpose
Hardware Cluster

• Jeanne Lehrter
• University of Tennessee
• May, 2002

2
Presentation Outline

• Introduction
• Background
• Grid-Based Software
• Our Work
• Conclusions
• Future Work

3
Introduction

• The need for grid computing
• SInRG research project
• NetSolve middleware by ICL
• Project Goals

4
Types of Hardware

• General purpose hardware can implement any
  function
• ASICs hardware that can implement only a
  specific application
• FPGAs reconfigurable hardware that can
  implement any function

5
FPGAs

• FPGAs offer reprogrammability
• Allows optimal logic design of each function to
  be implemented
• Hardware implementations offer acceleration over
  software implementations which are run on general
  purpose processors

6
Grid-Based Software

• NetSolve middleware
• Client-server-agent model
• Loosely-coupled heterogeneous environment
• No root or superuser privileges are required to
  run any part of NetSolve

7
NetSolve Agent

• Resource broker
• Maintains database of server characteristics and
  usage statistics
• Runs on Linux and UNIX

8
NetSolve Client

• Provides access to remote resources
• Interactive and programming interfaces
• NetSolve tools available to client
• Runs on Linux, UNIX, and Windows

9
NetSolve Server

• Offers resources to remote users
• Expandable capability
• Requires problem description file (PDF) for new
  functions to be added
• Contains a source code generator that creates
  service functions

10
Cluster for Advanced Machine Design

• Grid service cluster (GSC) in ECE
• Eleven SUN 220R Dual 450MHz UltraSPARC II
  processors running Solaris
• Eight Pentium III PCs running Linux with Xilinx
  Virtex-1000 chip attached
• A1000 RAID data storage unit
• 1 Gbit/s Foundry switch

11
Pilchard Environment for the FPGAs

• Developed at Chinese University in Hong Kong
• Plugs into 133MHz RAM DIMM slot and is an example
  of programmable active memory
• Pilchard is accessed through memory read/write
  operations
• Offers higher bandwidth and lower latency than
  other environments in which FPGA is placed in a
  PCI slot

12
Server Software Requirements

• NetSolve server software (and all software that
  is required by the NetSolve software, eg C
  compiler)
• Library of functions to be offered by the server
• PDFs that define the interfaces to these functions

13
Extra Software Requirements for ECE GSC

• Two types of functions are implemented in the ECE
  GSC
• Software version - runs on the PCs processor
• Hardware version - runs in the FPGA
• To implement the hardware version of the
  function, the software requirement of VHDL code
  is added

14
Implementing the Hardware Function

• The function is implemented in VHDL (or some
  other hardware description language)
• The VHDL code is then mapped onto the FPGA
  through the synthesis process
• Mapping decisions are based on constraints such
  as chip area, I/O pin counts, routing resources
  and topologies, partitioning, resource usage
  minimization

15
Implementing the Hardware Function Cont.

• Synthesis process transforms the VHDL description
  of a function to a configuration file which is a
  bit stream
• Configuration file defines how the FPGA is to be
  reprogrammed in order to implement the new
  desired functionality
• Hardware version of a function will copy the
  configuration file and function parameters onto
  the FPGA for processing

16
Server Example

• Consider this function interface to the software
  implementation of the Fast Fourier Transform
• void fft(char inputfile, charoutputfile, int
  size)
• Create a PDF that enables the NetSolve server to
  interface with this function by hand or by using
  PDF GUI generator

17
Software FFT PDF

• _at_PROBLEM fft
• _at_INCLUDE lt/avocado/homes/lehrter/my_netfuncs/libfp
  ga.hgt
• _at_LIB /avocado/homes/lehrter/my_netfuncs/libfpga.a
• _at_LANGUAGE C
• _at_MAJOR ROW
• _at_PATH /fpga/

18
Software FFT PDF Cont.

• _at_DESCRIPTION
• It performs an in-place decimation-in-frequency
  operation
• (Sande-Tukey FFT), leaving the output shuffled.
  Bit
• reversal is not performed.
• _at_INPUT 2
• _at_OBJECT FILE CHAR infile
• input file
• _at_OBJECT SCALAR I size
• pointer to size

19
Software FFT PDF Cont.

• _at_OUTPUT 1
• _at_OBJECT FILE CHAR outfile
• output file
• _at_CALLINGSEQUENCE
• _at_ARG I0
• _at_ARG O0
• _at_ARG I1
• _at_CODE
• fft(_at_I0_at_, _at_O0_at_,_at_I1_at_)
• _at_END_CODE

20
Hardware FFT

• Interface to hardware version of FFT is
• void ffta(char inputfile, char outputfile, int
  size)
• PDF for ffta is very similar to fft PDF except
• _at_PROBLEM ffta
• _at_CODE
• ffta(_at_I0_at_, _at_O0_at_, _at_I1_at_)
• _at_END_CODE

21
Server Configuration File

• _at_PROBLEMS
• ./problems/testing
• ./problems/qsort
• ./problems/mandelbrot
• ./problems/blas_subset
• ./problems/lapack_subset
• ./problems/fft
• ./problems/ffta

22
Client Code Example for fft

• include ltstdio.hgt
• include "netsolve.h"
• main(int argc, char argv)
• int info, result, a
• aatoi(argv3)
• info netsl("fft()", argv1, argv2, a)

23
Client Code for fft cont.

• if(info lt 0)
• netslerr(info)
• exit(0)
• if(info ! 0)
• printf("cannot solve\n")
• else
• printf("problem fft solved \n")

24
Client Code for ffta

• Client code for calling ffta is similar to fft
  client code except a different function is
  requested from NetSolve
• info netsl("ffta()", argv1, argv2, a)

25
Client Code Makefile

• NETSOLVE_ROOT /home/lehrter/NetSolve-1.4
• NETSOLVE_ARCH i686_pc_linux_gnu
• CC gcc
• INCDIR -I(NETSOLVE_ROOT)/include
• NETSOLVECLIB (NETSOLVE_ROOT)/lib/(NETSOLVE_ARC
  H)/libnetsolve.a
• CFLAGS (INCDIR) -g
• LIBS -lnsl
• test-fftns test-fftns.c
• (CC) (CFLAGS) -o _at_ test-fftns.c
  (NETSOLVECLIB) (LIBS)
• test-fftans test-fftans.c
• (CC) (CFLAGS) -o _at_ test-fftans.c
  (NETSOLVECLIB) (LIBS)

26
Running Client Code

• setenv NETSOLVE_AGENT utk1
• test-fftns infile outfile 512
• Initializing NetSolve...
• Initializing NetSolve Complete
• Sending Input to Server utk1
• Downloading Output from Server utk1
• problem fft solved
• test-fftans infile outfile 512

27
Behind the Scenes of the Client Code
VHDL code
Configuration file
Software and Hardware functions
PDFs, Libraries
28
Implementations

• Fast Fourier Transform (FFT)
• Data Encryption Standard algorithm (DES)
• Image Backprojection algorithm

29
Conclusions

• ECE GSC will offer hardware acceleration to
  general users
• Remote users can benefit from these online
  resources
• Resources are available through an efficient and
  easy-to-use interface

30
Future Work

• Determine how much speedup is achieved through
  hardware implementation of prototype functions
• Further increase the ECE GSC capabilities
• Find appropriate server benchmarking techniques
  for the ECE GSC hardware
• Create code that detects the current
  configuration loaded on the FPGA