Date: Mon, 17 Mar 2003 07:28:31 +0000 From: Antony T Curtis <antony.t.curtis@ntlworld.com> To: freebsd-java@FreeBSD.ORG Subject: JDK 1.3.1 performance Message-ID: <200303170728.33168.antony.t.curtis@ntlworld.com>
index | next in thread | raw e-mail
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Java benchmark using Volano client/server Results are scaled, bigger is better. classic 1.000 (1094 messages per second) +tya 1.107 +shujit 1.114 +OpenJIT 1.154 native 0.675 +tya 0.868 +shujit 0.833 +OpenJIT 0.847 HotSpot -Xint 0.983 client 1.427 server 0.746 jdk1.4 -Xint 0.896 client 1.301 server 1.234 These benchmarks should be important to people running networked java programs. Please note, I only ran the tests once, and both client and server are sitting on the same box. The JDK1.3.1 is with my own patches (includes some maybe bad fixes for - -native). I expected -native to have a much lower performance than -classic as FreeBSD#s userland threads have much heavier context switching than Java's green threads (which by nature, know about what it's running). What is surprising is that -server performs so badly whereas on 1.4, it performs reasonably. These results tell me that we need KSE's for native threading to be truely effective and that we should not implement Java's threading on top of a KSEified pthread but directly on FreeBSD-5's KSE api. I'm tempted to roll a different threading for native on FreeBSD 4 based upon rfork()... Just my $0.02 worth... - -- Antony T Curtis BSc Unix Analyst Programmer http://homepage.ntlworld.com/antony.t.curtis/ -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.2.1 (FreeBSD) iD8DBQE+dXkhql7dp2cddmIRAks7AJsFL37Oq9EhCi2dEZjueynkGSfhoQCff/7G dhgy8klgwXt873e3DEGJB00= =F5XB -----END PGP SIGNATURE----- To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-java" in the body of the messagehelp
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200303170728.33168.antony.t.curtis>