Re: Pgsql (and mysql) benchmark on T2000/Solaris and some

You usertime is way too high for T2000...

If you have a 6 core machine with 24 threads, it says all 24 threads are
reported as being busy with iostat output.

Best way to debug this is use

prstat -amL
(or if you are dumping it in a file prstat -amLc > prstat.txt)

and find the pids with high user cpu time and then use the usrcall.d on
few of those pids.

Also how many database connections do you have and what's the type of
query run by each connection?

-Jignesh

Arjen van der Meijden wrote:
> Hi Jignesh,
>
> The settings from that 'special T2000 dvd' differed from the recommended
> settings on the website you provided. But I don't see much difference in
> performance with any of the adjustments, it appears to be more or less
> the same.
>
> Here are a few iostat lines by the way:
>
> sd0 sd1 sd2 nfs1 cpu
> kps tps serv kps tps serv kps tps serv kps tps serv us sy wt id
> 7 1 12 958 50 35 0 0 7 0 0 0 13 1 0 85
> 0 0 0 2353 296 3 0 0 0 0 0 0 92 7 0 1
> 0 0 0 2062 326 2 0 0 0 0 0 0 93 7 0 0
> 1 1 1 1575 350 0 0 0 0 0 0 0 92 7 0 1
> 0 0 0 1628 362 0 0 0 0 0 0 0 92 8 0 1
>
> It appears to be doing a little less kps/tps on sd1 when I restore my
> own postgresql.conf-settings. (default wal/checkpoints, 20k buffers, 2k
> work mem).
>
> Is it possible to trace the stack's for semsys, like the memcpy-traces,
> or are those of no interest here?
>
> Best regards,
>
> Arjen
>
>
> On 16-5-2006 17:52, Jignesh K. Shah wrote:
>
>> Hi Arjen,
>>
>> Can you send me my colleagues's names in a private email?
>>
>> One of the drawbacks of the syscall.d script is relative timings and
>> hence if system CPU usage is very low, it gives the relative weightage
>> about what portion in that low is associated with what call.. So even
>> if you have say..1% system time.. it says that most of it was IO
>> related or semsys related. So iostat output with -c option to include
>> CPU times helps to put it in the right perspective.
>>
>>
>> Also do check the tunables mentioned and make sure they are set.
>>
>> Regards,
>> Jignesh
>>
>>
>> Arjen van der Meijden wrote:
>>
>>> Hi Jignesh,
>>>
>>> Jignesh K. Shah wrote:
>>>
>>>> Hi Arjen,
>>>>
>>>> Looking at your outputs...of syscall and usrcall it looks like
>>>>
>>>> * Spending too much time in semsys .... which means you have too
>>>> many connections and they are contending to get a lock.. which is
>>>> potentially the WAL log lock
>>>>
>>>> * llseek is high which means you can obviously gain a bit with the
>>>> right file system/files tuning by caching them right.
>>>>
>>>> Have you set the values for Solaris for T2000 tuned for Postgresql?
>>>
>>>
>>>
>>> Not particularly, we got a "special T2000 Solaris dvd" from your
>>> colleagues here in the Netherlands and installed that (actually one
>>> of your colleagues did). Doing so all the "better default"
>>> /etc/system-settings are supposed to be set. I haven't really checked
>>> that they are, since two of your colleagues have been working on it
>>> for the mysql-version of the benchmark and I assumed they'd have
>>> verified that.
>>>
>>>> Check out the tunables from the following URL
>>>>
>>>> http://www.sun.com/servers/coolthreads/tnb/applications_postgresql.jsp
>>>>
>>>> Try specially the /etc/system and postgresql.conf changes and see
>>>> if it changes/improves your performance.
>>>
>>>
>>>
>>> I will see that those tunables are verified to be set.
>>>
>>> I am a bit surprised though about your remarks, since they'd point at
>>> the I/O being in the way? But we only have about 600k/sec i/o
>>> according to vmstat. The database easily fits in memory.
>>> In total I logged about 500k queries of which only 70k where altering
>>> queries, of which almost all where inserts in log-tables which aren't
>>> actively read in this benchmark.
>>>
>>> But I'll give it a try.
>>>
>>> Best regards,
>>>
>>> Arjen
>>>
>>>>
>>>> Arjen van der Meijden wrote:
>>>>
>>>>> Hi List,
>>>>>
>>>>> In the past few weeks we have been developing a read-heavy
>>>>> mysql-benchmark to have an alternative take at
>>>>> cpu/platform-performance. Not really to have a look at how fast
>>>>> mysql can be.
>>>>>
>>>>> This benchmark runs on mysql 4.1.x, 5.0.x and 5.1.x and is modelled
>>>>> after our website's production database and the load generated on
>>>>> it is modelled after a simplified version of our visitor behaviour.
>>>>>
>>>>> Long story short, we think the test is a nice example of the
>>>>> relatively lightweight, read-heavy webapplications out there and
>>>>> therefore decided to have a go at postgresql as well.
>>>>> Of course the queries and indexes have been adjusted to (by our
>>>>> knowledge) best suit postgresql, while maintaining the same output
>>>>> to the application/interface layer. While the initial structure
>>>>> only got postgresql at about half the performance of mysql 4.1.x,
>>>>> the current version of our postgresql-benchmark has quite similar
>>>>> results to mysql 4.1.x, but both are quite a bit slower than 5.0.x
>>>>> (I think its about 30-40% faster).
>>>>>
>>>>> Since the results from those benchmarks are not yet public (they
>>>>> will be put together in a story at our website), I won't go into
>>>>> too much details about this benchmark.
>>>>>
>>>>> Currently we're having a look at a Sun T2000 and will be looking at
>>>>> will be looking at other machines as well in the future. We are
>>>>> running the sun-release of postgresql 8.1.3 on that T2000 now, but
>>>>> are looking at compiling the cvs-head version (for its
>>>>> index-root-cache) somewhere this week.
>>>>>
>>>>> My guess is there are a few people on this list who are interested
>>>>> in some dtrace results taken during our benchmarks on that T2000.
>>>>> Although my knowledge of both Solaris and Dtrace are very limited,
>>>>> I already took some samples of the system and user calls. I used
>>>>> Jignesh Shah's scripts for that:
>>>>> http://blogs.sun.com/roller/page/jkshah?entry=profiling_postgresql_using_dtrace_on
>>>>>
>>>>>
>>>>> You can find the samples here:
>>>>> http://achelois.tweakers.net/~acm/pgsql-t2000/syscall.log
>>>>> http://achelois.tweakers.net/~acm/pgsql-t2000/usrcall.log
>>>>>
>>>>> And I also did the memcpy-scripts, here:
>>>>> http://achelois.tweakers.net/~acm/pgsql-t2000/memcpysize.log
>>>>> http://achelois.tweakers.net/~acm/pgsql-t2000/memcpystack.log
>>>>> (this last log is 3.5MB)
>>>>>
>>>>> If anyone is interested in some more dtrace results, let me know
>>>>> (and tell me what commands to run ;-) ).
>>>>>
>>>>> Best regards,
>>>>>
>>>>> Arjen
>>>>>
>>>>> ---------------------------(end of
>>>>> broadcast)---------------------------
>>>>> TIP 3: Have you checked our extensive FAQ?
>>>>>
>>>>> http://www.postgresql.org/docs/faq
>>>>
>>>>
>>>>
>>>> ---------------------------(end of
>>>> broadcast)---------------------------
>>>> TIP 6: explain analyze is your friend
>>>>
>>>
>>> ---------------------------(end of broadcast)---------------------------
>>> TIP 2: Don't 'kill -9' the postmaster
>>
>>
>>