AV-Portal 3.23.2 (82e6d442014116effb30fa56eb6dcabdede8ee7f)

Benchmarking FreeBSD

Video in TIB AV-Portal: Benchmarking FreeBSD

Formal Metadata

Benchmarking FreeBSD
Benchmarking - what not to do and how to avoid it if possible
Title of Series
CC Attribution 3.0 Unported:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Release Date

Content Metadata

Subject Area
System optimization and tuning is tricky business. Benchmarking such systems is even more complicated as the number of things which can go wrong at doubles at the least. FreeBSD is best known for its killer features, but it also includes over 2500 sysctls, most of which can be tuned to do something interesting. This talk aims to give an overview of some of the more interesting things which can be tuned in FreeBSD, and advice on how to avoid the most common errors in benchmarking FreeBSD. Tuning a system heavily depends on hardware present in the system, so all tuning advice will necessary contain system-specific parts, but overall there is much to discuss when talking about optimizing specific systems: networking, storage, file systems, even the CPU scheduler. Networking is still very much dependant on the quality of the NIC and its driver, but large parts of the systems such as ZFS are pure software and can benefit from tweaks and tuning which slightly alter the behaviour of algorithms. Things get very complicated when comparing different hardware configurations, and even more when comparing different operating systems. Doing a good benchmark of two unrelated operating systems is tricky because it requires similar tune-ups to both systems. This talk will try to explain where the pitfalls are, and also present some field results.
Computer animation Information Operating system Value-added network
Pixel Email Freeware Trail Software developer Multiplication sign Software developer System administrator Electronic mailing list Sound effect Heat transfer Measurement Frame problem Benchmark Computer animation Bit rate Blog Blog System programming System programming Energy level
Number Product (business) Revision control Medical imaging Mathematics Benchmark Operator (mathematics) System programming Descriptive statistics World Wide Web Consortium Pairwise comparison Server (computing) Model theory Sound effect Port scanner Cartesian coordinate system Complete metric space Measurement Benchmark Computer animation Integrated development environment Computer hardware Order (biology) System programming Resultant Library (computing)
Pairwise comparison Enterprise architecture Presentation of a group Server (computing) Building Service (economics) Computer file Forcing (mathematics) Software developer Core dump Branch (computer science) Mereology Theory Neuroinformatik Revision control Database normalization Invariant (mathematics) Computer animation Internetworking Computer hardware Software Software testing Right angle Routing
Machine vision Regulärer Ausdruck <Textverarbeitung> Observational error Bit rate Mereology Approximation Value-added network Twitter Bit rate Average Different (Kate Ryan album) Natural number System programming Error message Position operator Social class Metropolitan area network Noise (electronics) Digitizing Bit Measurement Computer animation Personal digital assistant Hard disk drive output Resultant
Standard deviation Confidence interval Multiplication sign Numbering scheme Heat transfer Mereology Inference Different (Kate Ryan album) Single-precision floating-point format File system System programming Utility software Data structure Extension (kinesiology) Position operator Form (programming) Metropolitan area network Curve NP-hard Computer file Graph (mathematics) Expression Measurement Sequence Benchmark Backtracking Process (computing) Computer animation Oval Logic System programming Right angle
Noise (electronics) Slide rule Statistics NP-hard Information Flash memory Computer file Design by contract Bit Measurement Computer animation Personal digital assistant System programming File system output System programming
Area Metropolitan area network Computer animation Bridging (networking) Gastropod shell Heat transfer Entire function
Interior (topology) Multiplication sign Reflection (mathematics) Maxima and minima Design by contract Mereology Sequence Partition (number theory) Medical imaging Computer animation Different (Kate Ryan album) File system System programming Right angle Summierbarkeit MiniDisc Reading (process) Partition (number theory)
Computer file Interior (topology) View (database) Mereology Bit rate Different (Kate Ryan album) Operator (mathematics) File system System programming Information security Installable File System Metropolitan area network Enterprise architecture Octahedron Software developer Computer file Projective plane Attribute grammar Bit Division (mathematics) Database Limit (category theory) Sequence Computer animation Personal digital assistant Different (Kate Ryan album) System programming Resultant Writing Reading (process) Row (database)
Sensitivity analysis Group action Computer file Maxima and minima Branch (computer science) Mereology Medical imaging System programming Utility software Multimedia MiniDisc Installable File System Metropolitan area network Zoom lens Block (periodic table) Computer file Sound effect Automaton Benchmark Computer animation Personal digital assistant Network topology Configuration space Right angle Quicksort Spectrum (functional analysis) Writing
Metropolitan area network Zoom lens Maxima and minima Median Student's t-test Port scanner Average Mereology Benchmark Arithmetic mean Computer animation Different (Kate Ryan album) Personal digital assistant Different (Kate Ryan album) System programming Right angle
Metropolitan area network Thread (computing) Dataflow Computer file Variety (linguistics) Block (periodic table) Computer file Cellular automaton Electronic mailing list Maxima and minima Water vapor Directory service System call Exclusive or Computer animation Vector space Single-precision floating-point format Operator (mathematics) Block (periodic table) Reading (process) Social class
Computer file Divisor Multiplication sign Range (statistics) Branch (computer science) Calculus of variations Mereology Number Frequency Read-only memory Operator (mathematics) Core dump File system Software testing Partition (number theory) Stability theory Zoom lens Theory of relativity Graph (mathematics) Scaling (geometry) Dataflow Key (cryptography) Database Limit (category theory) Benchmark Open set Type theory Computer animation Personal digital assistant Order (biology) Universe (mathematics) Right angle
Server (computing) Statistics Presentation of a group Virtual machine Water vapor Heat transfer Mereology Perspective (visual) Benchmark Semiconductor memory Operator (mathematics) Core dump Square number File system System programming Integer Gamma function Logic gate Metropolitan area network Noise (electronics) Zoom lens Proper map Information Data storage device Database Limit (category theory) Measurement Benchmark Connected space Computer animation Personal digital assistant Musical ensemble Resultant Row (database)
Ocean current Server (computing) Concurrency (computer science) Service (economics) Consistency Virtual machine Client (computing) Field (computer science) Semiconductor memory Different (Kate Ryan album) Bridging (networking) File system System programming Metropolitan area network Graph (mathematics) Real number Exponentiation Ext functor Database Bit Line (geometry) Limit (category theory) Benchmark Connected space Computer animation Personal digital assistant Configuration space Right angle Remote procedure call Resultant Local ring
Point (geometry) Server (computing) Regulärer Ausdruck <Textverarbeitung> Service (economics) Datei-Server Proxy server Computer file Virtual machine Directory service Client (computing) Mereology Value-added network Number Twitter Benchmark Profil (magazine) Different (Kate Ryan album) Operator (mathematics) System programming Gamma function Information security World Wide Web Consortium Metropolitan area network Zoom lens Algorithm Email Server (computing) Computer file Electronic mailing list Attribute grammar Instance (computer science) Line (geometry) Benchmark Measurement Open set Connected space Arithmetic mean Computer animation Logic System programming Different (Kate Ryan album) Speech synthesis Configuration space Resultant Local ring
Dataflow Datei-Server Multiplication sign Physical law Line (geometry) Limit (category theory) Mereology Measurement User profile Inference Latent heat Computer animation Average Different (Kate Ryan album) Chain Hard disk drive Office suite Error message Resultant Local ring World Wide Web Consortium
Server (computing) Socket-Schnittstelle Computer file Dependent and independent variables Virtual machine Sheaf (mathematics) Water vapor Client (computing) Value-added network 2 (number) Cache (computing) Profil (magazine) Semiconductor memory Ring (mathematics) Software testing Office suite Data structure Socket-Schnittstelle Server (computing) Client (computing) Density of states Limit (category theory) Datei-Server Benchmark Message passing Computer animation Computer hardware Universe (mathematics) Resultant
Point (geometry) Group action Socket-Schnittstelle Server (computing) Observational error Direction (geometry) Execution unit Maxima and minima Control flow Streaming media Student's t-test Mereology Perspective (visual) Variance Neuroinformatik Number Wave packet Revision control Semiconductor memory Natural number Different (Kate Ryan album) Operator (mathematics) Entropie <Informationstheorie> System programming Multiplication Metropolitan area network Socket-Schnittstelle Noise (electronics) Zoom lens Model theory Message passing Process (computing) Computer animation Software Interrupt <Informatik> Block (periodic table) Spacetime
Game controller Implementation Computer file View (database) Compiler Drop (liquid) Mereology Revision control Benchmark Different (Kate Ryan album) Internetworking Befehlsprozessor Encryption System programming Utility software Implementation Gamma function Data compression Compilation album Area Metropolitan area network Algorithm Assembly language Shared memory Bit Port scanner Compiler Type theory Keilförmige Anordnung Kernel (computing) Computer animation Personal digital assistant Finite difference Software testing Data Encryption Standard Lipschitz-Stetigkeit Quicksort Musical ensemble Remote Access Service Library (computing)
Boss Corporation Curve Execution unit Projective plane Virtual machine Knot Benchmark Scalability Number Benchmark Computer animation Software Personal digital assistant Befehlsprozessor System programming Operating system Physical system Boss Corporation
Ocean current Point (geometry) Presentation of a group Linear regression Software developer Multiplication sign Software developer Perspective (visual) Value-added network Number Computer animation Bridging (networking) System programming Musical ensemble Analytic continuation
I think you're remember having numerous and I'm here to talk about benchmarking which is called the police and all everybody's excited to new benchmarks from out there everybody's trying to either the them or prove them wrong and something I did was they should be most about How to certain aspects of it or advanced users would probably know among so this is to know how to do and what not to do what we do information on the operating system and the thing have also
tried to the some new measurements of previously kind of you know extensive and useful for the years to come the last 1 last transfer done by scan rate those in and 8 this is the time frame of the center pixel release it to the and also address some the effects of being the mailing lists and on various blogs reach measured certain aspects of previously performance and features which are not really a professionally done and which are being so the picture for today is and I tried to I them on no them lesson Soseki rate so resembles
developers I mean the developers were interested in measuring performance of last resort of systems or subsystems they're working on a system of annotations system and this is the 1st we also find some useful material here and last year I want to go for fuel size represents is about how and when to avoid because sometimes you know you don't really need but generally need the best but
that means you have some purpose in mind I you want comparisons the summit something some other system you want to compare you have some other however you want to compare your application to some other applications it is mostly useless to perform better than prequential on itself so if you have a number that says you know that many megabytes per 2nd order of operations per 2nd and you don't hear anything very it just mostly most and you can compare it to a previous version of a system that's also well of you just have to be competitive system of the new Harbor but generally the African Union that fucking ministers have to compare it to something and the goal of a good
benchmark is estimated to be the about to make it useful product for other people but make description of image as library or as complete as possible so that people can be and see for themselves what you do is that would work on their own measurements comparable to yours so most of them are going to talk extensively or how the brain works and so I did were set up and what it is you actually arrived at the results repeatability is always a good thing even if you're going measure for only for yourself it means that the next week or next month or next year you will be able to create the same environment and it wish to check something such change for example is the new version of an application of our approach the effect of model performance I hope to
get some hi and however is but the question so all of this is all done 2 and computers so this idea I 1 you to and only a for who so if you without having to so for courses there working you get in this presentation this be a best friend and force certain rights you because they wanted to assess performance another theory redundancy like ability and now networking working was done on data with which has 2 parts 1 of which is connected to the Internet and 1 which is used to connect the better size of the service routes through simple wire so we don't switch the mean in and the in between
some notes about of versions I did both there is another on 1 which is the release version and the with the current pitches the development of the future of building that but this present then was for approximately 2 weeks ago so it's really recent of course the Department of certain all turned all witnesses invariances such things for comparison clinics and youth centers 6 . 3 because it is really probably the most used the so called enterprise version of Linux I was this girl to what branch money lost lost file bench and my own with server I will discuss it later and as preliminary
also that I talk about what and how to interpret at put the post-processed rhetoric results and this is a picture which is in every college textbook about think any kind of measurements so I will also repeated here average of measurements results the input was collected there's a or % precisely and you get accurate the measure the real the real of a on the as for the system you're trying to measure and at that precise they're all classes around basically the same river which should be that you want to read you to measure but these are usually is it'll be market clearing examples but it's
also related to to due to introduce measurement errors because domestic those which are uh influencing amendment in a way that you measure something else there was some interesting examples so basically you're thinking you're mentions 1 aspect of the system and the nature the mission something different and also random errors which are basically noise measurement noise which can be the trend come from old parts of the system you can basically I reduce its or just limited and no this textbook establish some kind of medium or average rate little bit that's also exaggerating their position so you can see in some benefits and incentives says you know how hard drive 400 and fails to this is the in this case failed . 5 6 to 7 megabytes per 2nd this is usually not because you cannot actually measure a part I want us to the performance that this position In most cases and during most cases the best you can do is measure you know what maybe the 1st significant digits so this will be probably be better expressed like 4 and it was interior you quote that you know really really really good this is just like in this case but in practice you will ever it's also useful of for
introduces the air about in the form of you know either on graphs and this smaller extensions in the graph below and above the real value was submitted something about the data basically as did universe or as an expression over the positions of your measurements confidence in that the measurements are really that's a precise or really repeatable in way it is you my
favorite example here is measuring public from performance because you get a lot of plants investments and and we just know that the data in a very or you just use a single job I've had J. about system on and just on some kind of this is not exactly the right it should be done because especially on mechanical Hargeisa so called spanning dressed such because you which differences in performance lean on sequential logic performances special right performance between the inner part of the driver and the outer part of the job that you view this scheme for utility which is part of every few years the support of the base system we can get a really nice the elicitation over how performance the 1st interval from the outside of the top but the inside of the job but it's so for example if you use if she just she passes them on such a drive and the measure prices in performance can be kind of utility and in really any kind of any kind of benchmark there at the time but it's really it's very much influence of being on which part of the drive your file system called places that the Soviets places the date that on the outside it gets what kind of performance the end and places in the inside you get another like that but that person performance curve this is this is the difference of approximately 15 100 megabytes per 2nd so really is significant so don't do persistent backtracking on hold jobs especially not on any device this is these here of different kind of problem and if you a lot of the internal structure of the politicization layer basically says that that your data which your right but that place drive can actually end up in some other places the job and you can be internally fragmented and it can be compressed and they can be no transfer in base you have no inference so basically pleasure is a whole other topic this is
also illustration of pure all POS system of this input based which is shown in the previous slide and it says it is is shows how well there's quite a lot of noise and the King contracts even those though this is this info so it doesn't pass your file system it talks directly to giant yeah that's Slashdot slavish use year or some other device on a specific but basically it says that there's still significant noise and this is noise between for example 209 with 2nd and 359 about respect it's significant it's the noise is also present in all cases the is the bit in the measurement is on the outside of the letters or on the inside of it is still present it is not use something you can so we use statistics to get an average how
well how many people have seen this
yeah basically the guys
shells the observed during this
entire area and bridges and the
the the guy shall so and his acoustic provision of
transfer to the giant carrot and
he sees a noticeable spike and lipids so really mechanical
device can be really really sense you get the invisible visible spot
this means that you know if you really
want to get scientific about it you
will have no need to go to the really much trouble to do it right fortunately and didn't always 14 ballpark for
example you always want to get values like you know g kind of megabytes per 2nd not
exactly 2 2 kinds of the 35 . 4 5 6 you really have to really think about your you
can expect so I would really want to delete image contracts for file system basically you want to a petition which is somewhere close to the outside of the and which is small enough so that the difference in performance is not noticeable between its beginning and it's an have this but this and has to be of reflection of size so you can run meaningful but for example it has to be at least 2 times larger than your available man because if you run a sequential read special right you need to ensure that the sum of cash but since all the that comes to pass system to make the and you care about of portion all file systems which is in your face approximately 8 per cent usually you have to do 2 times or 4 times the size of your artist petition and as an illustration of this couple side this is my prediction and on the right side is again that you saw in the whole John performance so when you look at just this petition that in the middle and on the outer middle and inner parts of this partition basically constant from was minus some not and of course we call giant is completely different the so but these are the
results I think the blue channel is not working on this project there should be a bar bias or in this case parliament wants and the red 1 is the 9th and the yellow 1 is reviewed them this result from 1 + + I'm not using the file system Bangladesh's sequential read and write rates and also of the view that you can see that the results are interesting and it's their capital if you know you can set much emphasis to the limits performance which is worse on I think and that from reading part the other way around what is interesting and this should be confirmed another best is that actually that's this given some development between 9 . 5 1 5 actually 1 police all previously and the gold was previously in such a way that basically the right performance is much better but also really performances a little bit about it means that some kind some kind of the world and this probably happen interior of Tennessee on maybe more frightening looking in some part of the system
a critical systems basically systems our kind of data you get records which are files stored on that these records are simple but the here is that it may be that the there is such a thing as security permissions here attributes but they also need to be chronically Oct exercise their freedom of writing soul this database really can get quite complex 0 this is notorious for carrying out really really complex layout and basically going more compared our muscles and compared to database than of division of and possibly the supports no other future should features slide it to be reliable that it's a need to support through operations and enterprise systems are also also complex bit different
next I would talk about the global bench for and it basically is a very small utility but the and written its goal is to create a tree smallish files the needs his group so files 1 is bonds sort of text files or who killed size for collection size and the other is like multimedia files or in this case image files which are in size but generally the sizes of people with the same spread across this spectrum between 2 kilobytes sensitive kilobytes the interesting thing is the 1 which is monitored traits hundreds hundreds of threats in the Kansas threats the people configuration stats Canada that that and all these threats are divided into several groups 1 group C is bundled creates files together to try and modifies fires that the dual trees 5 in this case the total mandatory files is the largest part so we can for example trade it trades beating pilots and the rest of it in the writing of treatment that they could those users automata names for some rights so for example modifies a block file and then it's renames it through right built on this is important because in previously much emphasis is given to the rightful king if you
seen in the olympics benchmarks you picture pictures like this so basically Linux performance and low branches old way on almost a million and actually the performance is where here so has thousand so it's puzzling so why would this be this always done new effects like it knows you get some rest you basically it comes down to to how they look at this system scandals reading and
writing what is interesting is that
there is some improvement between 9 and 10 reference and Detroit you can
use the where easy really convenient that would be present in previously this is Minister it is also part of the base system so we already have 2 benchmarks once this can fall and by means of a a lot to do with ways this can for 1 this means that it can be used to create quality than but I was was wondering what's the difference between the difference between what book 1 and then releases and it says that approximately 90 % a better performance is achieved in 10 but in in the real case approximately 17 % performance is achieved for the right for that means that
water investigation I will branch some of the it true it uses the bottle operations on the century or file and this something called the Viterbi so previously basically means that variety on single file block all other X I think if you work it's mature think this such a it useful to remember that the directories the themselves are also files so I right operations on the vectors themselves but also blocked I
did some physical tracing this you get along a long list of classes calls and it's clearly shows that there are literally multitude of threads during various operations opening reading writing closing and from all this data
that we can create a pretty picture which kind of characterizes the number and frequency of operation of what is shown in this x is the use of this 0 should be operation duration fights so there really very small and you order all of microseconds or where you give it a go at all rounded up to 0 . about 0 but you see it from this graph that low branch doesn't of but but you know you before we the relations are dominant not surprisingly there close operations good can the normal range you know uh range opens files modifies the more recent imposes them again and this was a really small number of rightful place basically this fair on here existing loan that read operations also they take the longest competitors the everything else which 18 Key indicates that there it should block by other types of cooperation but this names of also mitigation
I the other benchmark which is often used in previous years is positive because is subdatabase and in this case and it was the intervention which are just a part of was real SOS-based our military initialized the benchmark database with the scale factor of 4 thousand which means 100 million databases approximately 15 you but size it's facing around this is deliberate because of I'd like to test for example you can see in the universe exists without without Chilean benchmarking the car driver at the same time and the contributions of course stable billing core tests and this situation so what about at the beginning of so that the global limits and triggers potent in previously here this can access the same partition so the same partition the same file system is used for both Linux benchmarks and and better this is the provisions talk about at the beginning so that the the hundreds of which is just uh is just a small variations between nexus that exists performance on outside middle and church what this
real is a fairly modern database which means you which means it has what's going on In the back the most important thing is that right basically should do alright operations was the square in the data doesn't gold directly through that they can restore itself it goes is is written in the for and this data is then transferred from this the database be stored itself the problem is that this can hear it an expected fights so for example if you're unsure short benchmark you couldn't hold benchmark with the data residing in the record lost and not in the that storage itself and you can run another band which is started and then right in the middle with much this transfer occurs between the radical works and the proper storage and this introduces a large amount of noise in your results so better grants need to read these 5 minutes and minute there are there are recommendations for you longer kinds of our and it is true show that there was this girl server itself the the itself be started the between better because Aristotle was just felt was solvated adjusted from the water from the record books to with the 3rd itself so we get a more consistent more the precise measurement and also it's a certain the number out of the q the human posed as well is the part of the system which basically cleans up the In case of of integer records records and create in countries some statistics information which is also influence your results so they can be scatter right now this particular machine which is
the quad-core core is you know something about I've learned benchmarks reach a read-only which right and also the investments which are on the file system the England try on the dance itself so scary and on the memory for the system which is what I'm stressing gates of limits and the and the present is of previously of also some remote perspective being human connection to the database from the observed I will talk about some of the results and will be there
so on Linux I discovered that there someone no almost no difference being fully Karush be the based on text for injured on the bench and its current file system Memory system of previously using PFS is fast and then using your affairs which is fully loaded on our fuel for only a few on database cont the being lines are basically limits and the lines BSD the this this here is very surprising it is also very repeatable get in all configurations all a better performance I cannot utilize his bicycle contact exponent what was reported to have been but now is that on this particular machine I'd get lower performance so reduce the than off-limits except for the did here many of you can see it but there are 2 lines here the dark blue 1 and the like the 1 the that 1 is this file system and the light blue 1 is a memory policies on the because they that dark red use of these problems system you press and there are like children use he and the memory system there is noticeable difference this is these
are the right results Americans Linux various the of 1 and the few is occurring the even though this particular benchmark should really depend on the this performance apparently doesn't the opened system has that church influence on the book performance an this graph shows benchmark world remote access performance so what they did was they use the same services 1 was obviously around the database and the other was friend with client at database was always on memory file system and this is the distinction between local exons which has the full lights here the dark blue 1 and about 3 months and the middle the remote access for here which is the light blue 1 and the yellow 1 you can see that the remote access helps a little bit because the best itself it's so is securing as across so in separating the bridge reply and machine you free some resources on the server itself but this is also almost entirely by the fact that there a TCP connection between those sheets even though it's just that required between the 2 2 the same general conclusion to be drawn to produce the field has some work to do in performance improvement and because we can now separated the clients on the on the machine I think that this there is such a course with some kind of sketch issue in previously in the original case you care more than that and the server processes on the same machine example if approximately FIL concurrent clients Linux here's some kind of a problem in so my best guess is that consideration
however the get some more optimistic was these are the results done by for instance and also the think Jefferson on are usually of different machines the dispersion curves reported course because for secure think or speech and all 80 trends in called the 1 system and the results are much better also shot of creation have basically in this configuration leads previously is because even better than Linux so all this and hope yet uh the algorithms used in achieving our last abilities probably favor system assists with a large number of so something has to be done to maybe closer together meanings which is in this lection performance happened deal approximately 32 32 clients sorted through connections to the this is also suppose 9 point 2 and this is also a benchmark all local reference line and the local server on the same machine under mismatched I would
like to talk about followed bench this is the brush work done bird sunlight systems making that it has a large number of profiles 1 4 5 examples of service basically traits so list of all logic large files files and the other is the principle on which is smaller parts of the different benchmarks operations being done on these files on this unfortunately I think that this question has some problems running on a few years the I have some strange results with the weather later and I must assure that it's really a correct which and not sure that this results in a show you are obviously correct on the previous decide I've done local drive measurements on your face on this petition was talking about and also the 1st 3 and press for because there's some talk on the mailing lists about because performance difference between 1st reinforced previously so as to find out
what's it I this is the 1st approach and the results of the strange on the 1 side the care Linux this which . from here the result is larger than a hundred megabytes per 2nd which is obviously to cation average is right or actually would be all right if it weren't for previous years of which is this red line here which almost 111 per 2nd the designers themselves a scan tool for 100 50 megabits per 2nd please so I think that that use difference between the the Linux results and the resultant Fugazi results are due to benchmark problems so this is an example of maybe systemic systematic errors versus the problem that I don't and of history and inference for these are the 2 barrister here that like the 1 and that 1 the light blue 1 is a history in about dark blue 1 is an office for an even on limits and that those 4 is significantly smaller for some reason I was talking I was hoping to talk through a trick that is not in the conference stressed out the previous results our a law that links the new Linux results I really can't explain why will call this performance for this so long because this is much lower than ever itself and the longest chain and the NFS itself is again significantly lower than that of benefits from was mounted and on the
other hand we have the reportable which again shows review quirky results but the difference this time limits as law performance on the local hard drive measurements and previously has huge over performance the same provisions against it looks to me like the best itself is wrong for the most part that is itself seems to be the idea of AI is a part of is a Linux specific and there some problems supporting the previous or something this is possible but not really probable that such results with him so that this is just this just goes to say that you need to choose a best practice carefully I have an
interesting side result is also mention this is the same file bench benchmark on the file server profile and I've used from using as an officer to resource and limits of universe decline and so so that's the results of basically the same in which isn't always encourage last
year I talked about might look so it's a memory test or a similar comment give the specific thing is that the the trace of his epigenetics to call off very small fusion sections like bytes by messages being going back and forth between the client and server and it is given was started and also uses non-blocking and if you a benchmark on Unix sockets meaning that both water the server and the client around the same machine and talk about structural user made sockets you get easily intersection 4 seconds on mid-range shot the result my
favorite previously in this particular management but for the most part the itself was available on previously by me so I was a lot of tuning fork to we i also would not other this as a truly good better be and Linux entropy is that it's more popular curiosity within measurement error so the same performance is the technical 1 and the released is that the better is
interesting because because it illustrates that have used is this perspective is fairly most general and there are rarely finally locks so this a people's actions to be seems to not influence each other at multiple concurrent concurrent streams can be used to reduce the without performance problems it also shows that might work for example a nearly 500 thousand bikers provides a 2nd that the direction this cooperation so this is all a lot of really small piece of the messages so I went I suspect strong there of the difference between the Supreme or the there train wire Quebec they need a wire connecting servers and units means sockets user made sockets performance of process 1 million Jackson's a 2nd but this is also naturally expected gain on the other on 1 hand kept the whole into it it's generally where you can the quality natural networking and the other under the cancer care only memory spacing between processes so 1 thing that's all the people are trying to figure out is what a tune on previously to get acceptable performance and I would argue that especially when it's released then and in the not having to listen and from there is not much you can do or actually the system of computers itself pretty well because an increase in the general recently also a about space I have a base in lower space so there no altitude didn't have that some tutorials and other retirees increasing the net force of constant and assume that the the strip on the acoustic features and the size I haven't actually noticed some significant differences on my previous message so as there is some difference I'm not so sure this is because all of uh maybe noise I that this model that version find out but the point is that people tend to use our she pretty good I'm also trying to increase the number of interrupts per 2nd allow for 1 year and driver of this accused and if I increase the idea it's also a very small improvement made a less than 1 per cent of performance so unless you really trying to squeeze out some wasn't performance from emotions you don't have to touch the presumption that 4 . operation conducted makes users governance basically it the size and number of internal mental internal control structures it's been antigens 4 years of students for years old and for example for the last example the student will which is made from New York to everyone who who head run at the you or was this real or something else which is a lot of processes 14 is also going from the 9th of releases so my point is no you don't have to rely on manual tuning as much as used previously the I
didn't give any CPU-specific better and by that I mean I didn't do any encryption I know any kind of compression best box in such because for most part their actually unless you have a broken compiler or unless you care about different tuning in your compression libraries of different algorithms in your lips C and you're not going to find any significant or any useful difference it's utterance so most of the methods which are now available on the internet so uh noticeable difference in the compression performance for example of the use GC or some other leaves the utility of compressed which file on Linux and that it again previously and then show for example that this performance is little bit about the other thing is you performances with these for for example realized and how would the leaves he was compiled itself for 1 type of is the didn't use an assembler optimized part of the compression algorithm because there is used but in so colleagues he was compiled at see only according we should lower performance that then the several mice so this is the cost of drop in such manner if you want to you compare the quality of sleep C and D the implementation for example you can because of what your so for the I O is the I simple also something this but then short you know what the rest you're not restricted for example the kernel in this case the benchmarking algorithms in your lips you can also press your compiler there is the case sort of all this compiler fault this will change in my control
and just and just as an illustration you can find you can find practice is floating around in the band rare here for about 1 side GCC and recent versions of this year as well and you know the other and the share a ceiling compiler and stroll and for example this article that shows you what the difference between of the land and uses the what's this picture doesn't show is that there are other benefits children completely different areas so in this case that is mostly you can gamble virtually there's no problem knowing whether it's like this but you can't and all the learning ventricles are compiled so if you compile of example of compression library ordering kitchen library with different compilers and different operating system you kids it's all right you can do it but I just realized but you're not exactly which working for example the view that existing kernel you're benchmarking kings of the quality of work in
sometimes benchmarking and he's done in circumstances which you don't require indelible freezes that you can why you can do benchmarks not but there are also reasons that is working really it's not as useful or numbers of those useful as you think you can choose an operative system based on the features based on the community there a large number of so support software you can also let's say that is cheaper to buy another machine and guns the slightest lower operating system dental and I don't want conclusion that is also good to keep in mind is that they actually previously gets better to the years and it goes inconsistently it's more and more scalable during the case can revenge so that's and the text there are a few busy has started you scalability on its if you are is succeeded but that's all was not and you know then text set up the realistically probably aimed for having being scalable or being whether the Linux other containing going to use so consistently consistently controlling it is very
much a modern system going forward you need this maybe of curiosity unit if you don't know what's going quantum system in you planning or project for a new project your boss tells it to and finally we got yeah and and this is the curve the most commonly used is blue or actually published benchmarks on so during
development some some some of the other develop a sense of the plus administrative so previously and you're trying to get up some continual bridge parking going go on so maybe the next music can or something like that because some of the 2 numbers because some nice performance through current improvements over the time course the year off is
I this is it it's um the point of this presentation was mainly to point out some interesting facts about perspective itself about what to do and what not to do when they can count on predict system so I hope you read inflows a reasonably interesting presentation here and think you for the at the of