Java/JDBC slow compared to Perl/DBD-Oracle

Question

Hopefully the subject is inflammatory enough to get some responses :-)
I'm thinking about converting a database-intensive Perl application to Java, but it's fairly database intensive. Things look pretty straightforward, but Oracle's OCI JDBC driver seems to run at about 2/3 the speed of Perl with the DBD-Oracle driver.
For comparison, here's the Java version:
=== SelectTest.java
import java.sql.*;
import oracle.jdbc.driver.*;
public class SelectTest {
public static void main(String args[]) throws Exception {
int count = Integer.parseInt(args[0]);
Class.forName("oracle.jdbc.driver.OracleDriver");
Connection conn = DriverManager.getConnection(
"jdbc:oracle:oci8:@orcl.world",
"scott", "tiger");
String statement = "select * from dual";
OraclePreparedStatement stmt =
(OraclePreparedStatement) conn.prepareStatement(statement);
stmt.defineColumnType(1, Types.VARCHAR);
stmt.setRowPrefetch(1);
for (int i = 0; i < count; i++) {
ResultSet rset = stmt.executeQuery();
rset.next();
String s = rset.getString(1);
rset.close();
stmt.close();
=== End of SelectTest.java
and here's the Perl version:
=== select_test.pl
#!/usr/bin/perl -w
use strict;
use DBI;
my $dbh = DBI->connect('dbi:Oracle:dev.cag', 'cag', 'gac');
my $sth = $dbh->prepare('select * from dual');
for (1 .. $ARGV[0]) {
$sth->execute;
my @data = $sth->fetchrow;
$sth->finish;
=== End of select_test.pl
On my Linux box at home, 'time java SelectTest 1' gives me:
1.120u 0.140s 0:01.35 93.3% 0+0k 0+0io 4334pf+0w
and 'time java SelectTest 1000' gives me:
2.290u 0.370s 0:03.87 68.7% 0+0k 0+0io 4334pf+0w
So the OCI driver can execute about 400 execute/next's per second.
In comparison, 'time select_test.pl 1' gives:
0.550u 0.070s 0:00.74 83.7% 0+0k 0+0io 1638pf+0w
and 'time select_test.pl 1000' gives:
1.210u 0.220s 0:02.41 59.3% 0+0k 0+0io 1638pf+0w
So the Perl/DBD-Oracle runs at about 600 execute/fetch's per second.
The results are similar on our Sun 3500.
I'm at a loss to explain this. I'm explicitly setting the prefetch, using the Oracle defineColumnType extension, and preparing the statement outside the loop. Amazingly, the Perl program beats it with practically no optimization.
This is actually a best-case scenario for the JDBC driver. In the worst-case scenario, the same program running against a table with 50 columns runs about 6 times faster with Perl/DBD than with Java/JDBC. (Large records seem to be a particularly bad case for the JDBC driver).
These tests are actually highly important for our application, which frequently issues single-row selects.
I'm desperate to increase the JDBC performance. Every other test I've run shows Java running several times faster than Perl. Once I can get this ironed out, we should be all set.
For the record, the tests were run using JDK1.2 against classes111.zip. The results were similar using classes12.zip.
Does anyone have any suggestions? I'm hoping there's something obvious I'm missing that will speed up my code to acceptable levels.

Answer 1

ddankwerth,
Actually, we're using persistent connections with mod_perl, so I'm not anticipating a big speed boost there.
Do you have a code example where the thin driver outperforms the OCI driver? In my tests, it's at least 30% slower.
One thing I noticed after posting my question was that the total idle time for the JDBC code was about 50% higher than for the Perl code. This makes me think that the JDBC driver is making more calls to the DBMS than necessary.
Is Oracle actively tuning the JDBC drivers? I get the impression from the classes12 README that the development is focusing more on keeping up with the JDBC standard than with performance.
Does anyone know if there any plans to improve performance?