On 1/19/2015 1:46 PM, olli hauer wrote: > On 2015-01-19 20:18, Kurt Jaeger wrote: >> Hi! >> >>>>> Yes, i have. I've solved this problem by moving the build-jails of >>>>> poudriere to an memory disk. This make poudriere no longer io-bund and >>>>> incredibly fast. And solve this issue ;) >> >>>> How did you do this ? I want to try this myself 8-} >> >>> I've hacked poudriere to run within a jail. >> >> Aha, the .m mountpoint. My test host has 32 GB, so 20 GB should not be >> a problem. >> >> Testport: www/p5-Selenium-Remote-Driver on 10.1-amd64, 9.3-amd64 and >> 8.4-i386. >> >> Results: >> >> old: 00:05:43 >> new: 00:05:11 >> >> old: 00:01:56 >> new: 00:00:12 >> >> old: 00:02:11 >> new: 00:00:14 >> >> Nice! >> > > Hi Kurt, > > are you running PD also in a jail? > > If not PD can be tuned by setting MFSSIZE *or* USE_TMPFS in poudriere.conf. > > On my system I have good results with 8 concurrent builds and MFSSIZE=6G or > 'USE_TMPFS=all'. > Fine tuning can be done with an additional SSD (look at `systat -iostat' > during a build) > > poudriere.conf: > > # When building packages, a memory device can be used to speedup the build. > # Only one of MFSSIZE or USE_TMPFS is supported. TMPFS is generally faster > # and will expand to the needed amount of RAM. MFS is a bit slower, but is > # more mature and can have its memory usage capped. > > # If set WRKDIRPREFIX will be mdmfs of the given size (mM or gG) > #MFSSIZE=4G > > # Use tmpfs(5) > ... > # all - Run the entire build in memory, including builder jails. > USE_TMPFS=all >
Why do people pick MFS over TMPFS? I've found MFS/UFS significantly slower than TMPFS on FreeBSD 10+. I'm very inclined to remove MFS support from Poudriere as it is far less supported as TMPFS and not tested well. I suspect the reason is due to size constraint not being supported in the past. TMPFS_LIMIT can be used just as MFSSIZE can be. -- Regards, Bryan Drewery
signature.asc
Description: OpenPGP digital signature
