I agree that we were talking about Kannel performance itself, isolated from apps themselves (Alex's point #4). But regarding all of his other points, I agree with all of them, and without having mentioned them before, that was the type of things I was thinking of.
Your tests having thrown similar results on a few different servers doesn't mean that Kannel will have the same performance on any server, it's quite obvious that results can vary a lot depending on stuff like the mentioned by Alex. 2010/8/14 Nikos Balkanas <nbalka...@gmail.com>: > Sorry Alex, > > What you say is largely unsupported and unsubstantiated, unless you have > some numbers to back it up. I do. > > 1 & 2) Results asked and tests given are for typical conditions (normal fs, > spool) and a medium 64bit Ubuntu, kernel 2.6.31-22 box (dual-core Xeon > @3.4GHz). Interestingly, widely different OS/hardware give similar results, > ie a low end 64bit Solaris 10 box (dual-core AMD Opteron CPU @1218 uHz) is > only 15% slower than the previous Ubuntu box. > > 3) Indeed using a DB for DLRs has a major effect, effectively cutting MT > response to half, as shown in the tests. Choice of DB, will also have > some minor effect, too. Fakesmpp can evaluate different DBs, index or other > DB optimization choices (ie InnoDB vs MyIsam). > > As mentioned in the tests (check mail on fakesmpp submission) this is for > 64bit architecture. Nowdays, typical conditions is 64bit compilation. > > 4) I am only benchmarking kannel. Applications are custom made and not part > of kannel. > > 5) Nobody is trying to predict anything. You have an average figure of > kannel's response, build your system and run some benchmarks to see where > you fair. > > 6) Again, figures asked were for kannel. SMSc is not a kannel component. > > You really should read my mail about fakesmpp submission. > > Some useful conclusions that you can take home: > > 1) MOs, using default service, are ~30% faster than MTs using DLRs. > 2) Internal DLRs are almost twice as fast as DB DLRs. > 3) Throttling errors will slow kannel to a crawl. Kannel retries them every > 30". > > BR, > Nikos > > ----- Original Message ----- > From: Alejandro Guerrieri > To: Nikos Balkanas > Cc: Juan Nin ; users@kannel.org > Sent: Saturday, August 14, 2010 2:46 PM > Subject: Re: Kannel profermance > Your performance tests, while worth looking as a general reference, are > probably near to useless as benchmarks for other people's platforms. > 1. Variations in hardware could alter your figures greatly. Differences in > CPU, memory and/or HDD speeds, as well as using RAID and/or SATA/SCSI/Fiber > interfaces (among many other factors) obviously will alter the results, and > it's _very_ difficult to predict the amount of variation, since it depends > greatly on how do you really use Kannel. > 2. Variations at the OS level also play a role: Using different Unix > flavors, different disk partitioning types, ramdisks and a whole lot of > other stuff that could be configured in a million different ways will also > alter the results, and in a very hard to predict fashion. > 3. Variations in Kannel configuration: the type of store, either using a DB > or not for DLR's (and the type of DB engine), either if the DB engine and > the application layer are local or remote and another thousand possible > factors. Also if you compiled it as 32 or 64 bit and what kind of compiler > optimizations you did. > 4. Last but not least, the application layer: yo don't specify what you're > doing with your messages when they're received, and here's when your > benchmarks become unrealistic: The response time of the applications is KEY > and can pose a severe bottleneck to Kannel, thus affecting the "raw > performance" in very unpredictable ways: An application that takes time to > respond will hold threads open, thus incrementing the "iowait" and "load" of > the Kannel box accordingly. That degrades a great deal performance. > So, since it's near impossible to predict all those factors and the > application level can play a key role into determining the final > performance, I wouldn't take your (nor anyone else's) performance benchmarks > as a strict reference of what Kannel is capable of regarding performance in > my particular platform. > I guess a proper statement could be "my Kannel setup running on OS A is > capable of handling B MO/s and C MT/s using this configuration {...}" but > again, external factors like application's speed, smsc throttling and a > thousand more factors could make very similar setups to perform in very > dissimilar ways. > A possible approach to provide a more relevant benchmark would be to > standarize a series of tests with some fixed variables (specially at the > application layer). That would provide at least a measure to compare against > between installs, but again, it wouldn't be able to predict anyone's > real-life scenario at all, as regular computer benchmarks are unable to > predict how fast your computer will be for the work you'll use it for. > In other words, don't take other people's benchmarks very seriously to > predict what your system will be capable of. > Regards, > Alex > 2010/8/14 Nikos Balkanas <nbalka...@gmail.com> >> >> Nope. The average between various servers that have been tested are >> approximately the ones cited. Standard deviation between optimized >> servers/OS is not that large (+/-50). This includes various >> conditions/architectures, each one with each own results. I just posted the >> most common ones. No optimized filesystem was used anywhere (spool, ext3 on >> linux, ufs on Solaris). Tests are very reproducible and discriminating with >> respect to DB, and possibly filesystem used for storage (not tested yet). >> >> If in doubt run your own tests. >> >> Nikos >> ----- Original Message ----- From: "Juan Nin" <jua...@gmail.com> >> To: <users@kannel.org> >> Sent: Saturday, August 14, 2010 8:23 AM >> Subject: Re: Kannel profermance >> >> >>> So you're saying that on any server and/or architecture the results >>> will be the same? >>> Doesn't seem very reasonable... >>> >>> >>> 2010/8/14 Nikos Balkanas <nbalka...@gmail.com>: >>>> >>>> Actually you have missed a couple of more emails. On fakesmpp submission >>>> I >>>> also posted results from a low-end Solaris 10 64bit box. Very similar to >>>> the >>>> results posted from the linux server. The averages seem pretty solid. >>>> So, >>>> contrary to your beliefs, it is not giving out the wrong impression. >>>> >>>> BR, >>>> Nikos >>>> ----- Original Message ----- From: "Juan Nin" <jua...@gmail.com> >>>> To: <users@kannel.org> >>>> Sent: Saturday, August 14, 2010 4:33 AM >>>> Subject: Re: Kannel profermance >>>> >>>> >>>>> Ok, just saw on another thread where you got those values from, but >>>>> again, that's very system specific. >>>>> >>>>> I guess you point was just to say that Kannel was not the issue for >>>>> his bottleneck, but saying "It can handle ~1000 MO/s, 750 MT/s >>>>> (internal DLRs) or 450 MT/s (DB DLRs)" may give the wrong impression >>>>> that that's the maximum it can handle, or that it can always handle >>>>> that load, while on low end servers it may not. >>>>> >>>>> So saying something like that in that way can confuse new users, IMHO. >>>>> >>>>> >>>>> On Fri, Aug 13, 2010 at 10:26 PM, Juan Nin <jua...@gmail.com> wrote: >>>>>> >>>>>> 2010/8/13 Nikos Balkanas <nbalka...@gmail.com>: >>>>>>> >>>>>>> It is unlikely that kannel is your bottleneck. It can handle ~1000 >>>>>>> MO/s, >>>>>>> 750 >>>>>>> MT/s (internal DLRs) or 450 MT/s (DB DLRs). >>>>>> >>>>>> >>>>>> Just curious to know where do you get those values from... >>>>>> What Kannel supports depends on your hardware and architecture >>>>>> >>>>> >>>>> >>>>> >>>>> -- >>>>> Juan Nin >>>>> 3Cinteractive / Mobilizing Great Brands >>>>> http://www.3cinteractive.com >>>>> >>>> >>>> >>> >>> >>> >>> -- >>> Juan Nin >>> 3Cinteractive / Mobilizing Great Brands >>> http://www.3cinteractive.com >>> >> >> > > -- Juan Nin 3Cinteractive / Mobilizing Great Brands http://www.3cinteractive.com