Tags: tuning


  08:42:00 pm, by The Dreamer   , 1176 words  
Categories: Software, Computer, Storage, FreeBSD, CFEngine

Another weekend seems to be slipping away on me....

And, its the same time suck....cacti.

Last weekend got away from me, because I to make another attempt to improve cacti performance. I had tried adding 3 more devices to it, and that sent it over the limit.

I tried the boost plugin....but it didn't help, and only made things more complicated and failure prone. Evidently, updating rrd files is not a constraint on my cacti server. Probably because of running on an SSD.

I made another stab at getting the percona monitoring scripts to actually work under script server, but that failed. I suspect the scripts aren't reentrant, because of their use of global variables and relying on 'exit' to cleanup things it allocates or opens.

I had blown some previous weekend when I had tried to build the most recent version of hiphop to maybe compile the scripts, but after all the work in figuring out how to compile the latest 2.0x version...it would SEGV, just as the older lang/hiphop-php did after resolving the problem of building with the current boost (a template had changed to need a static method, meaning old code won't link with newer boost libraries without a definition of this.) And, this is beyond what I have in my wheelhouse to try to fix.

During the week, I had come across some more articles on tuning FreeBSD, namely a discussion of kern.hz for desktop vs servers. Where it being 1000 by default is good for desktops, but the historical setting of 100 being what to use for servers. Though IIRC, ubuntu uses 250 HZ for desktops and 100 HZ for servers, it also doesn't do preemption in its server kernel along with other changes (wonder if some of those would apply to FreeBSD?) Though modern kernels have been moving to be tickless. Which I thought was in for FreeBSD 9, though the more correct term is dynamic tick mode...and which is more about not doing unnecessary work when things are idle. Which isn't the case with 'cbox'. So, perhaps, fiddling with kern.hz and other sysctls might still be relevant. Though haven't really found anything detailed/complete on what would apply to my situation.

So, I thought I would give kern.hz=100 a shot.

At first it seemed to make a difference....no improvement in how long to complete a poll, but the load was lower. Until I realized that a service had failed to start after reboot. I had only run the rc script by hand, I hadn't tested it in a reboot situation. And, its not an rc script....it was used to be a single line in rc.local that worked on ubuntu and FreeBSD (except on one of the Ubuntu systems it results in a ton of zombie processes, so making it an init.d script that I could call restart on happened.

So, I spent quite a lot of time reworking it into what will hopefully be an accept rc script. One thing I had changed was that instead of using a pipe ('|') which was causing the process after the pipe to respawn and turn the previous process into a zombie each time the log file was rotated and "tail -F" announced the switch. And, this was while I was moving the service to FreeBSD (and management under cfengine 3.)

Though looking at my cacti graphs later....while the service had failed to start after reboot, it turned out to have been running for sometime, until I had broken it completely in trying to rc-ify the init script. Will, duh....I had cfengine set to promise that the process was running, and it had repaired that it hadn't started after the reboot.

Another thing I had done with I had init-ified the startup of this service, was I switched from using pipe ('|') to using a fifo, which addressed the respawning and zombie problem and eliminated the original reason to have an init.d script....

While the init.d script had worked on FreeBSD...it was just starting the two processes with '&' on the end then exiting. FreeBSD's rc subroutines do a bit more than that. So things weren't working. The problem was that even though I was using daemon instead of '&', so that daemon would capture the pid and make a pidfile. seems daemon wants the process it manages to be fully working before it'll detach. But, the process is blocked until there's a sink on the other end of the fifo. (does sink fit was the name for the fifo's reader?) I first wonder if I could just flip the two around, but I suspect starting the read process first would be just as blocked until the write process is started. So, I cheated by doing a prestart of the writing process and only tracking the reading process.

Though it took a bit more work to get the 'status' action to work....eventually found I needed to define 'interpreter' since the reading process is a perl script. And, the check_pidfile does more than just check to see if there's a process at the pid, but that its the right process. And, it distinguishes between arg0 and the rest.

Pretty slick...guess I need to do a more thorough reading of the various FreeBSD handbooks, etc. Of course, it has been 13+ years between when I first played with FreeBSD to its take over of my life now.

As for the tuning....it had made a small difference, but no improvement on cacti system stats. Basically the load average fluctuates a bit more and the CPU utilization seems to be a bit lower...though it could because the 4 lines of the cacti graph aren't so close to each other now.

Meanwhile...I noticed that one of the block rules in my firewall had a much higher count than I would expect, so I think I was about to get logging configured to see what that's about.....(which I was working on when I remembered that I hadn't rebooted after making the kern.hz change to /boot/loader.conf yesterday...the commit also picked up files that I had touched while working on moving the one remaining application on 'box', though that may get delayed to another weekend....perhaps the 4 day one coming up.)

I had set cf-execd's schedule to be really infrequent (3 times an hour), because I was doing a lot of testing and cf-agent collisions are messy....messier than they were in cfengine 2 (in 2 it usually just failed to connect and aborted, in 3 it would keep trying and splatter bits and pieces everywhere....which is bad when there are parts using single copy nirvana. resulting in services getting less specific configs, until the next run.

But, I sort of brought back dynamic bundle sequences.... but key off of "from_cfexecd", so I can test my new promise with less problems of colliding with established promises. Though there are other areas where things still get messy.... need to clean up some of the promises I had based on how things were done at work, so that the promises are more standalone.

Kind of weird using my home cfengine 3 setup, and other admin activities, as the means to break the bad habits I had picked up at work....

Now instead of subjecting some poor random forum to a long rambling thought, I will try to consolidate those things into this blog where they can be more easily ignored profess to be collected thoughts from my mind.

Latest Poopli Updaters -- http://lkc.me/poop


There are 20 years 3 months 28 days 15 hours 8 minutes and 19 seconds until the end of time.
And, it has been 4 years 8 months 29 days 22 hours 54 minutes and 37 seconds since The Doctor saved us all from the end of the World!


September 2017
Mon Tue Wed Thu Fri Sat Sun
 << <   > >>
        1 2 3
4 5 6 7 8 9 10
11 12 13 14 15 16 17
18 19 20 21 22 23 24
25 26 27 28 29 30  


  XML Feeds

Who's Online?

  • Guest Users: 1
This seal is issued to lawrencechen.net by StopTheHacker Inc.
powered by b2evolution

hosted by
Green Web Hosting! This site hosted by DreamHost.

monitored by
Monitored by eXternalTest
SiteUptime Web Site Monitoring Service
website uptime