There is a 4% regression in JetStream benchmark across all platforms in Tiger b05. When comparing Xprof output between b04 and b05 we see 11% more time spent in StringBuffer append and 7% more time spent in StringBuffer toString methods. Looking at the IO workspace archives, the regression was introduced at

/net/teal/a/builds/tiger/linux-i586/2003/063a

We know this since the regression is not present in the previous 063 archive build.

xprof of 063
     Compiled + native Method
  9.5% 54 + 0 java.io.BufferedReader.read
  8.6% 49 + 0 java.io.BufferedOutputStream.write
  8.2% 11 + 36 java.lang.String.getChars
  6.7% 38 + 0 java.io.BufferedInputStream.read
  6.7% 21 + 17 java.lang.String.<init>
  6.0% 32 + 2 java.lang.StringBuffer.append
  4.0% 9 + 14 java.lang.StringBuffer.append
  4.0% 23 + 0 sun.nio.cs.ISO_8859_1$Encoder.encodeArrayLoop
  3.7% 21 + 0 sun.nio.cs.ISO_8859_1$Decoder.decodeArrayLoop
  3.5% 19 + 1 WriteTextTest.writeFile3
  3.2% 18 + 0 java.io.StreamTokenizer.nextToken
  2.8% 16 + 0 java.lang.Integer.toString
  2.6% 15 + 0 java.io.BufferedReader.readLine
  2.1% 12 + 0 java.lang.String.<init>
  2.1% 12 + 0 sun.nio.cs.StreamEncoder.write
  1.9% 11 + 0 java.io.StreamTokenizer.read
  1.8% 10 + 0 java.lang.Integer.getChars
  1.8% 9 + 1 java.lang.Integer.appendTo
  1.4% 8 + 0 java.lang.StringBuffer.length
  1.2% 7 + 0 java.nio.CharBuffer.wrap
  1.1% 6 + 0 java.io.PrintWriter.write
  1.1% 6 + 0 java.nio.charset.CharsetEncoder.encode
  0.9% 5 + 0 java.io.OutputStreamWriter.write
  0.9% 5 + 0 java.nio.Buffer.position
  0.7% 4 + 0 java.lang.ThreadLocal$ThreadLocalMap.get
 91.9% 454 + 71 Total compiled (including elided)

xprof of 063a
 9.4% 33 + 25 java.lang.StringBuffer.append
  8.4% 52 + 0 java.io.BufferedInputStream.read
  7.8% 48 + 0 java.io.BufferedOutputStream.write
  7.3% 45 + 0 java.io.BufferedReader.read
  6.2% 26 + 12 java.lang.StringBuffer.toString
  5.5% 34 + 0 sun.nio.cs.ISO_8859_1$Decoder.decodeArrayLoop
  5.0% 31 + 0 sun.nio.cs.ISO_8859_1$Encoder.encodeArrayLoop
  4.9% 30 + 0 java.io.StreamTokenizer.nextToken
  4.7% 15 + 14 java.lang.String.<init>
  4.2% 14 + 12 java.lang.StringBuffer.append
  3.9% 10 + 14 java.lang.String.getChars
  2.8% 14 + 3 java.lang.Integer.toString
  2.6% 14 + 2 WriteTextTest.writeFile3
  2.3% 14 + 0 sun.nio.cs.StreamEncoder.write
  2.3% 14 + 0 java.io.BufferedReader.readLine
  1.8% 4 + 7 java.lang.Integer.appendTo
  1.6% 10 + 0 java.io.StreamTokenizer.read
  1.5% 9 + 0 java.nio.CharBuffer.wrap
  1.3% 2 + 6 java.lang.Integer.toString
  1.3% 8 + 0 java.lang.Integer.getChars
  1.1% 7 + 0 java.io.OutputStreamWriter.write
  1.1% 5 + 2 java.io.BufferedWriter.write
  1.0% 6 + 0 CopyFileTest.copyFilesInDir
  0.8% 5 + 0 java.io.PrintWriter.write
  0.8% 5 + 0 java.nio.Buffer.limit
 92.9% 476 + 97 Total compiled (including elided)


We believe the 2% regression in Specjbb on Solaris is also caused by changes to StringBuffer toString (although not necessarily caused by build 063). Getting Xrunhprof:cpu=samples output and running a diff on the output, we have three different places where StringBuffer toString time increases.

Before: 243/34506 (0.704225%)
After: 452/34383 (1.3146%)
        java.lang.StringBuffer.toString(StringBuffer.java)
        spec.jbb.Company.getCustomerByLastName(Company.java)
        spec.jbb.PaymentTransaction.process(PaymentTransaction.java)
        spec.jbb.TransactionManager.go(TransactionManager.java)
---
Before: 211/34506 (0.611488%)
After: 419/34383 (1.21863%)
        java.lang.StringBuffer.toString(StringBuffer.java)
        spec.jbb.JBButil.create_random_last_name(JBButil.java)
        spec.jbb.JBButil.choose_random_last_name(JBButil.java)
        spec.jbb.PaymentTransaction.init(PaymentTransaction.java)
---

Before: 188/34506 (0.544833%)
After: 354/34383 (1.02958%)
        java.lang.StringBuffer.toString(StringBuffer.java)
        spec.jbb.History.buildData(History.java)
        spec.jbb.PaymentTransaction.process(PaymentTransaction.java)
        spec.jbb.TransactionManager.go(TransactionManager.java)


###@###.### 2003-05-07

Volano too {(d32, d64) sparc, win32, linux, solx86} is impacted by these changes.


###@###.### 2003-05-08