A running description of activity related to DragonFly BSD.
Matthew Dillon did some more performance tuning for DragonFly. I’ll just pull a paragraph from the commit message, since that will have more impact than anything I say:
Improves fork/exec concurrency on monster of static binaries from 14200/sec to 55000/sec+. For dynamic binaries improve from around 2500/sec to 9000/sec or so (48 cores fork/exec’ing different dynamic binaries). For the same dynamic binary it’s more around 5000/sec or so.
“monster” is a 48-core machine used for testing.
Branched, not released. The release should happen in two weeks. One major bug has been squished, and remember the upgrade process from 3.4 to 3.6 is a little different from normal.
DragonFly developer Francois Tigeot was interviewed on linuxfr.org. As you can probably guess from the names, it’s a French site, but don’t let that stop you if you’re an Anglophone.
The venerable (from 1979!) program, lpr, has been superseded by CUPS in many installations. Francois Tigeot suggested removing it, but it’s still directly usable in specific situations and easier to just shift out of the way. It’s staying, but it’s interesting to see how it still gets used.
Update: Predrag Punosevac has descriptions of the various tools involved.
Multiple transmission queues are possible for the mxge(4) driver; I’m mentioning it because Sepherosa Ziehau’s post about this describes the exact tunables to configure this.
I’m planning to branch DragonFly 3.6 this weekend. The actual release will come 2 weeks later. (Ignore what I wrote about a dports installer/image.)
Matthew Dillon wrote a roundup post summarizing all the changes he’s made to DragonFly to improve SMP performance in the last few weeks. He’s removed almost all contention from DragonFly. This means better performance, scaling upward depending on the number of processors.
‘monster’, the system that builds all 20,000 items in dports, can complete the run in 15 hours. Compare this to the 2 weeks it used to take me to build the 12,000 packages in pkgsrc. This is admittedly on different hardware and different packaging systems, but it gives a sense of the scale of the improvement.
Joris Giovannangeli, who worked on porting Capsicum to DragonFly for Summer of Code 2013, is continuing his work. He’s posted a detailed note on how to do capability management in a new way, with it retaining compatibility with FreeBSD’s capsicum implementation.
The ‘poweroff’ command, the equivalent of ‘halt -p’, has been added based on a suggestion from Robin Hahling.
Matthew Dillon has gone after reducing contention and improving SMP performance as vigorously as possible, using dports builds on a 48-processor machine as a test. The machine’s building more than 1000 packages an hour, last I saw on IRC.
John Marino has updated ldns and drill to version 1.6.16.
Chris Turner was curious about clock skew when running under a VM, and Sepherosa Ziehau very kindly explained the different types of internal clock for DragonFly, and how to control them – a topic I’ve never understood deeply.
Matthew Dillon was using poudriere, the dports build tool, on a 48-core system. Poudriere was building all 20,000+ dports, so the machine was quite busy. He decided to get rid of as much contention as possible, and he’s listed all the ways DragonFly’s been streamlined by these efforts. We need to revisit some of our previous benchmarks…
There is a search plugin for Mozilla that searches DragonFly man pages. (Thanks Samuel Greear)
I stole Sepherosa Ziehau’s email subject for the title of this post, because that’s exactly what has happened. Gigabit networking cards under DragonFly will perform very well under extreme load – all of them.
As a followup to news that the git feed of pkgsrc through dragonflybsd.org is not being updated, Max Herrgard wrote out how to fetch pkgsrc via CVS, or tarball, or another git feed. CVS is still the ‘official’ way.
Matthew Dillon’s been working to make huge parallel software builds (i.e. dports) go a bit faster, so watch out. This only affects you if you are running DragonFly 3.5, of course.
The pkgsrc repository in git for DragonFly is currently frozen. This is because many people have switched over to dports, and also because it’s a lot of work to keep it functional. If you do want to pull newer pkgsrc material, use cvs and grab it from a NetBSD server.
As the message notes, don’t go switching to DragonFly-current right now, cause there’s a lot of new material in there and it may not be quite safe. (There’s an ABI change that will require all new builds of your ports, for instance.)
If you’re watching for a certain event, dmesg(8) on DragonFly now has a -f flag that will display new output as it’s logged, sort of like ‘tail -f’.
The Radeon KMS driver from FreeBSD has been imported to DragonFly by Francois Tigeot. It still has problems with ttm, but don’t let that stop you from taking advantage of it.