JavaMemoryModel: Some notes on the DB benchmark

From: Bill Pugh (pugh@cs.umd.edu)
Date: Tue May 23 2000 - 13:01:33 EDT


I've been preparing some analysis of synchronization overhead in
benchmarks for the JavaOne talk I'm doing with Doug.

Here are some interesting results on the spec db benchmark, which is
single threaded.

The following gives execution time (in seconds) for different
versions of the the Spec db benchmark.

Original - using the original benchmark
ArrayList - convert the benchmark to use ArrayList rather than Vector
New Algorithm - replace the hand coded shell sort with a call to the
        built in merge sort
Both - both of the above two transformations.

Sync - Using the standard rt.jar
Unsync - Using a version of rt.jar in which synchronization has been removed
         from Hashtable, Vector, Stack, BufferedInputStream and
         BufferedOutputStream.

Original Sync 44.6
Original NoSync 36.3
ArrayList NoSync 34.1
ArrayList Sync 34.1
New Alg Sync 19.5
Both Sync 15.5
Both NoSync 14.5
New Alg NoSync 14.3

The lesson from this: yes, there is some synchronization overhead, but it is
        dwarfed by the stupidity overhead

By the way, the version that uses merge sort reports "Incorrect"
results. The reason for this is that the merge sort is stable, while
the shell sort is not. In truth, you'd prefer a stable sort. But it
does give you different results.

        Bill
-------------------------------
JavaMemoryModel mailing list - http://www.cs.umd.edu/~pugh/java/memoryModel



This archive was generated by hypermail 2b29 : Thu Oct 13 2005 - 07:00:25 EDT