New Phoronix performance benchmarks between some Linuxes and *BSDs
dennis berger
db at bsdsystems.de
Tue May 28 15:04:49 UTC 2013
Sorry, I missed the "variable file sizes" part.
So forget about my post.
Am 28.05.2013 um 16:27 schrieb dennis berger:
> Hi,
> for me it's unknown what 100 TPS means in that particular case. But this doesn't make sense at all and I don't see such a low number in the postmark output here.
>
> I think I get around 4690+-435 IOPS with 95% confidence.
>
> Guest and the actual test system is FreeBSD9.1/64bit inside of Virtualbox.
> Host system is MacOSX on 4year old macbook
> Storage is VDI file backed on a SSD (OCZ vortex 2) with a 2gb ZFS pool
>
> When you I postmark with 25K transactions I get an output like this. (http://fsbench.filesystems.org/bench/postmark-1_5.c)
>
> pm>run
> Creating files...Done
> Performing transactions..........Done
> Deleting files...Done
> Time:
> 6 seconds total
> 5 seconds of transactions (5000 per second)
>
> Files:
> 13067 created (2177 per second)
> Creation alone: 500 files (500 per second)
> Mixed with transactions: 12567 files (2513 per second)
> 12420 read (2484 per second)
> 12469 appended (2493 per second)
> 13067 deleted (2177 per second)
> Deletion alone: 634 files (634 per second)
> Mixed with transactions: 12433 files (2486 per second)
>
> Data:
> 80.71 megabytes read (13.45 megabytes per second)
> 84.59 megabytes written (14.10 megabytes per second)
>
> I did this 100 times on my notebook and summed up this.
>
> root at freedb:/pool/nase # ministat -n *.txt
> x alltransactions.txt
> + appended-no.txt
> * created-no.txt
> % deleted-no.txt
> # reed-no.txt
> N Min Max Median Avg Stddev
> x 100 3571 5000 5000 4690.25 435.65125
> + 100 1781 2493 2493 2338.84 216.8531
> * 100 1633 2613 2613 2396.59 256.53752
> % 100 1633 2613 2613 2396.59 256.53752
> # 100 1774 2484 2484 2330.22 216.3084
>
>
> When I check "zpool iostat 1" I see
>
> root at freedb:/pool/nase # zpool iostat 1
> capacity operations bandwidth
> pool alloc free read write read write
> ---------- ----- ----- ----- ----- ----- -----
> pool 10.6M 1.97G 0 8 28 312K
> ---------- ----- ----- ----- ----- ----- -----
> pool 10.6M 1.97G 0 33 0 4.09M
> ---------- ----- ----- ----- ----- ----- -----
> pool 10.6M 1.97G 0 0 0 0
> ---------- ----- ----- ----- ----- ----- -----
> pool 10.6M 1.97G 0 0 0 0
> ---------- ----- ----- ----- ----- ----- -----
> pool 10.6M 1.97G 0 0 0 0
> ---------- ----- ----- ----- ----- ----- -----
> pool 19.6M 1.97G 0 89 0 4.52M
> ---------- ----- ----- ----- ----- ----- -----
>
>
> around 30-90 TPS bursts.
>
> Did they counted this instead?
>
>
> -dennis
>
>
>
>
>
>
> Am 28.05.2013 um 15:02 schrieb Paul Pathiakis:
>
>> Outperform at "out of the box" testing. ;-)
>>
>> So, if I have a "desktop" distro like PCBSD, the only thing of relevance is putting up my own web server???? (Yes, the benchmark showed PCBSD seriously kicking butt with Apache on static pages.... but why would I care on a desktop OS?)
>>
>> Personally, I found the whole thing lacking coherency and relevancy on just about anything.
>>
>> Don't get me wrong, I do like the fact that this was done. However, there are compiler differences (It was noted many times that CLANG was used and it may have been a detriment but it doesn't go into the how or why.) and other issues.
>>
>> There was a benchmark on PostGreSQL, but I didn't see any *BSD results?
>>
>> Transactions to a disk? Does this measure the "bundling" effect of the "groups of transactions" of ZFS? That's a whole lot less transactions that are sent to disk. (Does anyone know any place where this can be found? That is, how does the whole "bundling of disk I/O" go from writing to memory, locking those writes, then sending all the info in one shot to the disk? This helps: http://blog.delphix.com/ahl/2012/zfs-fundamentals-transaction-groups/ )
>>
>> I was working at a company that had the intention of doing "electronic asset ingestion and tagging". Basically, take any thing moved to the front end web servers, copy it to disk, replicate it to other machines, etc... (maybe not in that order) The whole system was java based.
>>
>> This was 3 years ago. I believe I was using Debian V4 (it had just come out.... I don't recall the names etch, etc) and I took a single machine and rebuilt it 12 times: OpenSuSe with ext2, ext3, xfs. Debian with ext2, ext3, xfs. CentOS with ext2, ext3, xfs. FreeBSD 8.1 with ZFS, UFS2 w/ SU.
>>
>> Well, the numbers came in and this was all done on the same HP 180 1u server rebuilt that many times. I withheld the FBSD results as the development was done on Debian and people were "Linux inclined". The requisite was for 15000 tpm per machine for I/O. Linux could only get to 3500. People were pissed and they were looking at 5 years and $20m in time and development. That's when I put the FBSD results in front of them..... 75,200 tpm. Now, this was THEIR measurements and THEIR benchmarks (The Engineering team). The machine was doing nothing but running flat out on a horrible method of using directory structure to organize the asset tags... (yeah, ugly) However, ZFS almost didn't care compared to a traditional filesystem.
>>
>> So, what it comes down do is simple.... you can benchmark anything you want with various "authoritative" benchmarks, but in the end, your benchmark on your data set (aka real world in your world) is the only thing that matters.
>>
>> BTW, what happened in the situation I described? Despite, a huge cost savings and incredible performance.... "We have to use Debian as we never put any type of automation in place that would allow us to be able to move from one OS to another"... Yeah, I guess a Systems Architect (like me) is something that people tend to overlook. System automation to allow nimble transitions like that are totally overlooked.
>>
>> Benchmarks are "nice". However, tuning and understanding the underlying tech and what's it's good for is priceless. Knowing there are memory management issues, scheduling issues, certain types of I/O on certain FS that cause it to sing or sob, these are the things that will make someone invaluable. No one should be a tech bigot. The mantra should be: "The best tech for the situation". No one should care if it's BSD, Linux, or Windoze if it's what works best in the situation.
>>
>> P
>>
>> PS - When I see how many people are clueless about how much tech is ripped off from BSD to make other vendors' products just work and then they slap at BSD.... it's pretty bad. GPLv3? Thank you... there are so many people going to a "no GPL products in house" policy that there is a steady increase in BSD and ZFS. I can only hope GPLv4 becomes "If you use our stuff, we own all the machines and code that our stuff coexists on" :-)
>>
>>
>>
>>
>>
>>
>> ________________________________
>> From: Adrian Chadd <adrian at freebsd.org>
>> To: O. Hartmann <ohartman at zedat.fu-berlin.de>
>> Cc: freebsd-performance at freebsd.org
>> Sent: Tuesday, May 28, 2013 5:03 AM
>> Subject: Re: New Phoronix performance benchmarks between some Linuxes and *BSDs
>>
>>
>> outperform at what?
>>
>>
>>
>> adrian
>>
>> On 28 May 2013 00:08, O. Hartmann <ohartman at zedat.fu-berlin.de> wrote:
>>> Phoronix has emitted another of its "famous" performance tests
>>> comparing different flavours of Linux (their obvious favorite OS):
>>>
>>> http://www.phoronix.com/scan.php?page=article&item=bsd_linux_8way&num=1
>>>
>>> It is "impressive, too, to see that PHORONIX did not benchmark the
>>> gaming performance - this is done exclusively on the Linux
>>> distributions, I guess in the lack of suitable graphics cards at
>>> Phronix (although it should be possible to compare the nVidia BLOB
>>> performance between each system).
>>>
>>> Although I'm not much impressed by the way the benchmarks are
>>> orchestrated, Phoronix is the only platform known to me providing those
>>> from time to time benchmarks on most recent available operating systems.
>>>
>>> Also, the bad performance of ZFS compared to to UFS2 seems to have a
>>> very harsh impact on systems were that memory- and performance-hog ZFS
>>> isn't really needed.
>>>
>>> Surprised and really disappointing (especially for me personally) is
>>> the worse performance of the Rodinia benchmark on the BSDs, for what I
>>> try to have deeper look inside to understand the circumstances of the
>>> setups and what this scientific benchmark is supposed to do and
>>> measure.
>>>
>>> But the overall conclusion shown on Phoronix is that what I see at our
>>> department which utilizes some Linux flavours, Ubuntu 12.01 or Suse and
>>> in a majority CentOS (older versions), which all outperform the several
>>> FreeBSd servers I maintain (FreeBSD 9.1-STABLE and FreeBSD
>>> 10.0-CURRENT, so to end software compared to some older Linux kernels).
>>> _______________________________________________
>>> freebsd-performance at freebsd.org mailing list
>>> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
>>> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
>> _______________________________________________
>> freebsd-performance at freebsd.org mailing list
>> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
>> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
>> _______________________________________________
>> freebsd-performance at freebsd.org mailing list
>> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
>> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
>
>
> _______________________________________________
> freebsd-performance at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-performance
> To unsubscribe, send any mail to "freebsd-performance-unsubscribe at freebsd.org"
Dipl.-Inform. (FH)
Dennis Berger
email: db at bsdsystems.de
mobile: +491791231509
fon: +494054001817
More information about the freebsd-performance
mailing list