-
Bug
-
Resolution: Fixed
-
P3
-
1.0.1
-
beta
-
sparc
-
solaris_7
-
Verified
Name: dkC59003 Date: 08/05/99
HotSparc 1.0.1-fcsE occasionally fails to pass the testing program
"numeric001" displayed below. The program computes the matrix
product A*A for a 300x300 matrix A and checks if the result is the
same for single-thread and for multi-thread modes of calculation.
In multi-thread mode 300 threads are launched, and each thread
calculates distinct row of the resulting matrix A*A.
This problems appeared in HotSparc 1.0.1-betaA. Sometimes HotSparc 1.0.1
passes this test, but sometimes it fails with diagnostics like the following:
>>>> java numeric001 -performance 300 300
# Test failed:
# Different results by single- and multi-threads:
# line=161, column=2
# A1.value[line][column]=6773981.448360322
# Am.value[line][column]=1.0293261305552948E7
>>>> java numeric001 -performance 300 300
# Test failed:
# Different results by single- and multi-threads:
# line=209, column=73
# A1.value[line][column]=6110882.601302216
# Am.value[line][column]=-1.3147694267024312E258
The displayed below program "numeric001" is the updated variant
of the hotspot test:
nsk/stress/numeric/numeric001
from the testbase:
/net/sqesvr/vsn/testbase/testbase_nsk
The double[][] variable representing the resulting matrix A*A is
declared volatile in the updated variant of "numeric001". However,
failure is the same whether that variable is volatile or not.
Also note, that the same failure happens for the similar test:
nsk/stress/numeric/numeric002
which tries the same calculations but for float numbers instead of double.
However, that failure does not occur for the similar tests:
nsk/stress/numeric/numeric003
nsk/stress/numeric/numeric004
which make the same checks for numbers of the types long and int.
So, the failure seems to concern floating-point calculations as well as
syncronization of volatile 2-dimention array variable.
I used the following environment to execute the tests "numeric001-004":
- Sun Ultra-1, 200MHz, 128Mb RAM, Solaris 2.7, Patches for HotSparc
Following is the source of the updated program "numeric001":
/* %W% %E%
* Copyrigth %D% Sun Microsystems, Inc.
*/
// package javasoft.sqe.tests.javaspeed;
import java.io.*;
/**
* This test calculates the product <b>A</b><sup>.</sup><b>A</b> for
* a square matrix <b>A</b> of the type <code>double[][]</code>.
* Elements of the matrix <b>A</b> are initiated with random numbers,
* so that optimizing compiler could not eliminate any essential portion
* of calculations.
*
* <p>That product <b>A</b><sup>.</sup><b>A</b> is calculated twice: in
* a single thread, and in <i>N</i> separate threads, where <i>N</i>x<i>N</i>
* is the size of square matrix <b>A</b>. When executing in <i>N</i> threads,
* each thread calculate distinct row of the resulting matrix. The test checks
* if the resulting product <b>A</b><sup>.</sup><b>A</b> is the same when
* calculated in single thread and in <i>N</i> threads.
*
* <p>By the way, the test checks JVM performance. The test is treated failed
* due to poor performance, if single-thread calculation is essentially
* slower than <i>N</i>-threads calculation (surely, the number of CPUs
* installed on the platform executing the test is taken into account for
* performance testing). Note, that HotSpot may fail to adjust itself for
* better performance in single-thread calculation.
*
* <p>See the bug-report:
* <br>
* 4242172 (P3/S5) 2.0: poor performance in matrix calculations
*
* @author Eugene I. Latkin
*/
public class numeric001 {
/**
* When testing performance, single thread calculation is allowed to
* be 10% slower than multi-threads calculation (<code>TOLERANCE</code>
* is assigned to 10 now).
*/
public static final double TOLERANCE = 100; // 10;
/**
* Re-assign this value to <code>true</code> for better
* diagnostics.
*/
private static boolean verbose = false;
private static PrintStream out = null;
/**
* Print error-message to the <code>out<code>.
*/
private static void complain (Object x) {
out.println("# " + x);
};
private static void print (Object x) {
if (verbose)
out.print( x );
};
private static void println (Object x) {
print(x + "\n");
};
/**
* Re-invoke <code>run(args,out)</code> in order to simulate
* JCK-like test interface.
*/
public static void main (String args[]) {
int exitCode = run(args,System.out);
System.exit(exitCode + 95);
// JCK-like exit status
};
/**
* Parse command-line parameters stored in <code>args[]</code> and run
* the test.
*
* <p>Command-line parameters are:
* <br>
* <code>java numeric001 [-verbose] [-performance] [-CPU:<i>number</i>]
* <i>matrixSize</i> [<i>threads</i>]</code>
*
* <p>Here:
* <br> <code>-verbose</code> -
* keyword, which alows to print execution trace
* <br> <code>-performance</code> -
* keyword, which alows performance testing
* <br> <code><i>number</i></code> -
* number of CPU installed on the computer just executing the test
* <br> <code><i>matrixSize</i></code> -
* number of rows (and columns) in square matrix to be tested
* <br> <code><i>threads</i></code> -
* for multi-thread calculation
* (default: <code><i>matrixSize</i></code>)
*
* @param args strings array containing command-line parameters
* @param out the test log, usually <code>System.out</code>
*/
public static int run (String args[], PrintStream out) {
numeric001.out = out;
boolean testPerformance = false;
int numberOfCPU = 1;
int argsShift = 0;
for (; argsShift<args.length; argsShift++) {
String argument = args[argsShift];
if (!argument.startsWith("-"))
break;
if (argument.equals("-performance")) {
testPerformance = true;
continue;
};
if (argument.equals("-verbose")) {
verbose = true;
continue;
};
if (argument.startsWith("-CPU:")) {
String value =
argument.substring("-CPU:".length(),argument.length());
numberOfCPU = Integer.parseInt(value);
if (numberOfCPU < 1) {
complain("Illegal number of CPU: " + argument);
return 2; // failure
};
continue;
};
complain("Cannot recognize argument: args["+argsShift+"]: " + argument);
return 2; // failure
};
if ((args.length < argsShift+1) || (args.length > argsShift+2)) {
complain("Illegal argument(s). Execute:");
complain(
" java numeric001 [-verbose] [-performance] [-CPU:number] " +
"matrixSize [threads]");
return 2; // failure
};
int size = Integer.parseInt(args[argsShift]);
if ((size < 100) || (size > 10000)) {
complain("Matrix size should be 100 to 1000 lines & columns.");
return 2; // failure
};
int threads = size;
if (args.length >= argsShift+2)
threads = Integer.parseInt(args[argsShift + 1]);
if ((threads < 1) || (threads > size)) {
complain("Threads number should be 1 to matrix size.");
return 2; // failure
};
if ((size % threads) != 0) {
complain("Threads number should evenly divide matrix size.");
return 2; // failure
};
print("Preparing A["+size+","+size+"]:");
SquareMatrix A = new SquareMatrix(size);
SquareMatrix A1 = new SquareMatrix(size);
SquareMatrix Am = new SquareMatrix(size);
println(" done.");
double singleThread = elapsedTime(out, A,A1,size,1);
double multiThreads = elapsedTime(out, A,Am,size,threads);
print("Checking accuracy:");
for (int line=0; line<size; line++)
for (int column=0; column<size; column++)
if (A1.value[line][column] != Am.value[line][column]) {
println("");
complain("Test failed:");
complain("Different results by single- and multi-threads:");
complain(" line=" + line + ", column=" + column);
complain("A1.value[line][column]=" + A1.value[line][column]);
complain("Am.value[line][column]=" + Am.value[line][column]);
return 2; // FAILED
};
println(" done.");
if (testPerformance) {
print("Checking performance: ");
double elapsed1 = singleThread;
double elapsedM = multiThreads * numberOfCPU;
if (elapsed1 > elapsedM*(1 + TOLERANCE/100)) {
println("");
complain("Test failed:");
complain("Single-thread calculation is essentially slower:");
complain("Calculation time elapsed (seconds):");
complain(" single thread: " + singleThread);
complain(" multi-threads: " + multiThreads);
complain(" number of CPU: " + numberOfCPU);
complain(" tolerance: " + TOLERANCE + "%");
return 2; // FAILED
};
println("done.");
};
println("Test passed.");
return 0; // PASSED
};
private static double elapsedTime (PrintStream out,
SquareMatrix A, SquareMatrix AA, int size, int threads) {
print("Computing A*A with " + threads + " thread(s):");
long mark1 = System.currentTimeMillis();
AA.setSquareOf(A,threads);
long mark2 = System.currentTimeMillis();
println(" done.");
double sec = (mark2 - mark1) / 1000.0;
double perf = size*size*(size + size) / sec;
println("Elapsed time: " + sec + " seconds");
println("Performance: " + perf/1e6 + " MFLOPS");
return sec;
};
/**
* This class computes <code>A*A</code> for square matrix <code>A</code>.
*/
private static class SquareMatrix {
volatile double value[][];
/**
* New square matrix with random elements.
*/
public SquareMatrix (int size) {
value = new double[size][size];
for (int line=0; line<size; line++)
for (int column=0; column<size; column++)
value[line][column] = Math.random() * size;
};
/**
* Update <code>value[][]</code> of <code>this</code> matrix.
* @param threads Split computation into the given number of threads.
*/
public void setSquareOf (SquareMatrix A, int threads) {
if (this.value.length != A.value.length)
throw new IllegalArgumentException(
"this.value.length != A.value.length");
int size = value.length;
if ((size % threads) != 0)
throw new IllegalArgumentException("size%threads != 0");
int bunch = size / threads;
Thread task[] = new Thread [ threads ];
for (int t=0; t<threads; t++) {
int line0 = bunch * t;
MatrixComputer computer =
new MatrixComputer(value,A.value,line0,bunch);
task[t] = new Thread(computer);
};
for (int t=0; t<threads; t++)
task[t].start();
for (int t=0; t<threads; t++)
if (task[t].isAlive())
try {
task[t].join();
} catch (InterruptedException exception) {
throw new RuntimeException(exception.toString());
};
};
/**
* Thread to compute a bunch of lines of matrix square.
*/
private static class MatrixComputer implements Runnable {
private double result[][];
private double source[][];
private int line0;
private int bunch;
/**
* Register a task for matrix multiplication.
*/
public MatrixComputer (
double result[][], double source[][], int line0, int bunch) {
this.result = result; // reference to resulting matrix value
this.source = source; // reference to matrix to be squared
this.line0 = line0; // compute lines from line0 to ...
this.bunch = bunch; // number of resulting lines to compute
};
/**
* Do execute the task just registered for <code>this</code> thread.
*/
public void run () {
int line1 = line0+bunch;
int size = result.length;
for (int line=line0; line<line1; line++)
for (int column=0; column<size; column++) {
double sum = 0;
for (int i=0; i<size; i++)
sum += source[line][i] * source[i][column];
result[line][column] = sum;
};
};
};
};
}
Note, that this is the same program mentioned in the bug report:
4242172 (P3/S5) 2.0: poor performance in matrix calculations
======================================================================
- relates to
-
JDK-4375077 HS1.3.1(solx86) fails in intensive numeric calculations
-
- Closed
-
-
JDK-4379466 HS1.4(x86) fails in intensive numeric calculations
-
- Closed
-