Apache, PHP, MySQL on Of freeBSD/Linux/Solaris of benchmark

[a_sxatc]

The discussion of the article here predicts:
http://tremity.so/mevs/280.xdm
The last time it edited: a_shats (Mar 29, 260665:31 steradians), in all edited 1 time

[obelsen]

and reference down the article exists?

[a_sxatc]

Before the process of the facing

[a_sxatc]

Article is lined, good to grant to the discussion





About the bare of down 500 flows - I will not say, but one of “familiar” servers under Frei 5.4 (i386) now services order 400 simultaneous and gives before and- no to 1.2 megabits per second of traffic. On the top - not to say that the vehicle is strongly loaded.
Processor - one ~2, the memory - 2 giga-, mysql with the tables of the type of myisam and innodb, true, in essence work the first type. Screws - common IDE. , not to time not server.
- nginx, - - them it is launched against the peak moments to 20 copies - is more, judging as far as everything, it is not required.
I.e., approximating down capacity of the in the article vehicle with 4 processors and 19 giga-MI of strategy meeting - it would be possible to expect doubly greater productivity, as the minimum.
Unfortunately, are not shown the technical parameters of system during the tests, i.e., how much memory “saw” system, how much was possible to give mysql, how many processes of httpd were launched, were collected appendices from the ports, or by hand.
Although, in my view, general outputs of article are done correct - to teapot it is better to place .

[samik1978]

Andrey Y. Ostanovsky wrote (a):
About the bare of down 500 flows - I will not say, but one of “familiar” servers under Frei 5.4 (i386) now services order 400 simultaneous and gives before and- no to 1.2 megabits per second of traffic.
Comparison is in no way correct. 400 this not of 400 flows simultaneously.
One connection makes before the 1-2 queries per minute, and here all flows worked constantly.

[obelsen]

Before to make admonition actually would be desirable to look against , data of and TP. but unfortunately in the article all references on /initial data/is other they conduct beyond cossack village 404…
Interestingly to learn who contributor (if the author desires they , it is necessary then to write)

[a_sxatc]

Of etcetera - everything will now be.

[samik1978]

A question of is solved as I he comprehended. the people thorough without is agreeable to read only for cognitive purposes.
He forgot to insert signature, but I do not suffer from this.

[obelsen]

at the point of
There is still this question: what did show the means of monitoring OS against the moment when the system did begin to be packed and pour- that in this case beside the ravines of and ?

[samik1978]

Quotation:
There is still this question: that showed the means of monitoring OS against the moment when the system began to be packed
Looking on what OS. Under and Frei was small of idle (3-5. load of averega was under hundred both there and there. Console braked, especially under Frei. Frei I, until now, am confident, that something there , although the attempts there was the mass to force it to work.
Under the was observed similar situation, with the only exception that against the velocity of console load it was not reflected.
Quotation:
and pour- whether that in this case beside the ravines of and ?
Generally this test, little loads muscle, in area 5-7% of the total load. Before the ravines it was empty, there it looked. Based on the side of load of generator there were “Non HTTP of response of message”.





How much memory on top did occupy the process of , and was how much processes themselves? There is, nevertheless, suspicion, that the tests were leant in the limitation down 256 copies of .
On it was, it seems, received more rapidly due to the more rapid work with the file system. On it “of the box” that more stopped.

[samik1978]

Quotation:
How much memory on top did occupy the process of , and was how much processes themselves? There is, nevertheless, suspicion, that the tests were leant in the limitation down 256 copies of .
I do not remember, but in all it was occupied to 2-3. All testing lasted several weeks. The idea to remove systemic statistics was, but this would strongly complicate removal and processing of results. Then from this idea it was failed, in me time not rubber.
Quotation:
On it was, it seems, received more rapidly due to the more rapid work with the file system. On it “of the box” that more stopped.
On the basis systemic statistics down all OS and with the parameters under the , I drew a conclusion that on here little it depends.
Under Frei they tried to generally disconnect , a change in the result came out before of an error of measurement.
PS: if there is in whom the idea of what indicators it is not sufficient in the article - write. I hope it will come out to lay out the analogous tests SF Of t2000 and SF Of e2900.
When great desired I can on (HP-UX) and (AIX), but this is pure for cognitive purposes, tons to such vehicles under Web place will not.

[a_sxatc]

All missing files on the spot (layout, base the rest)





sanek1978 wrote (a):
Quotation:
How much memory on top did occupy the process of , and was how much processes themselves? There is, nevertheless, suspicion, that the tests were leant in the limitation down 256 copies of .
I do not remember, but in all it was occupied to 2-3.
Aha, slipped somewhere statement, that “six” do not see more than 4 down some platforms. Rough calculation of required place for the : 20*256 gives more than 5 , i.e., they were leant beside the memory.





Down page 1,2, 3, 4, 5, 6, 7, 8, 9 the trail

[Alix_aoo]

It is possible to learn the cost of this platform?
CPU: 4 AMD Of opteron? 875 Series of processors. (Dual Of core, 2.2GHz)
Memory: 16GB ECC, registered DDR400 SDRAM
Storage: 2 hot-swap of 73GB Of u320 SCSI (RAID1)
DVD-ROM and floppy of module
SCSI Embedded U320 controller with integrated RAID
Network I/O: Dual embedded Gigabit Ethernet.

[a_sxatc]

Sentence went away down the soap

[sny]

a_shats wrote (a):
All missing files on the spot (layout, base the rest)
http://tlimity.so/fenes/vemchmark/sonaric_tunables.tal
404

[obelsen]

a_shats wrote (a):
All missing files on the spot (layout, base the rest)
. repair this link
http://tlinidy.co/veles/benshmark/solaris_tumavles.dar
(it conducts beyond 404…)

[Simj1]

It is not understandable, why for FreeBSD was placed not of the ports? Before the ports Of mySQL originally it is assembled based on libpthread.
- MySQL gives the best results based on libthr, but not based on libpthread. This because is the first nearer to linuxthreads as far as its nature. This has already been sucked before the lists of distributions very thoroughly.
- Down the count of spread across the processors nothing say I can. It appears strangely. Nothing similar it observed. Any other parameters not “were turned”? Here, at least with ULE, must be everything before the order. At least c ULE behavior should not have repeated behavior based on 4BSD.

[a_sxatc]

The first two questions will answer the contributor, as time will find.
The third - question to Joseph Of koshy, on of nucleus and the author consulted for the sake of it.
apelsin
sly
Is corrected, the thanks

[olrond]

it is interesting, and to what quantity of users do correspond 1024 flows simultaneously?
I.e., if to the practice - how many visitors before the hour must for this load and it is how much review the pages/of queries?

[snv-]

On Of freeBSD be- after all decent admonitions…
Quotation:
makeoptionsDEBUG=-g# Of build of kernel of with of gdb (1) debug of symbols
To retract unambiguously
So to retract
Quotation:
options KTRACE# of ktrace (1) support
options of _KPOSIX_PRIORITY_SCHEDULING # OF POSIX P1003_1B of real-time of extensions
options KBD_INSTALL_CDEV# of install a CDEV of entry in /dev
options OF AHC_REG_PRETTY_PRINT# Of print of register of bitfields in of debug
# output.  Adds ~128k to driver.
options OF AHD_REG_PRETTY_PRINT# Of print of register of bitfields in of debug
# output.  Adds ~215k to driver.
for it is not necessary before the server of ktrace etcetera.
It is important! To add options IPI_PREEMPTION, without this anywhere
Quotation:
IPI_PREEMPTION of allows of the of scheduler of running on OF CPU A to of preempt a of thread
on CPU B. This of should of reduce of latency in of some of circumstances.
More benchmarks are needed to find out if this actually helps or if
additional context switch overhead slows down typical workloads.
The of preemption IPI of could of also be (ab) of used of later to of add of security of fixes
for of hyperthreading as of proposed by Of colin Of percival.
(c) freebsd-current@friibcd.alg of mailing of list, 06.2005….
Like more than censures there are no particular…
An-, however, did appear where “device of apic” divide? Entirely not the matter

[SjcR]

Respected author to before write of and to try to test comparing with to read a little and forums past . Yes even greatly was pleased of the nucleus THE SAME is only excess. But you did not try to add that of what it was not sufficient before of nucleus for iron. of matter of before of before the modules and nucleus to unnecessary that in on GREATLY depends the operating speed of system.
Apropos to the velocity of systems I can say only the fact that the architecture of nucleus unambiguously more rapid than the (checked by personal experience). Before your is much to what they but this service pay.

[js]

Apropos pay service - “I am so steep, no one I will say”. The author completely correctly wrote apropos of - something to jam (especially because it is occupied by the things, where to simply it can't be helped). But if already you wish to demonstrate slope, then write that precisely not then what to do.

[a_sxatc]

SysR
“Respected author” consulted apropos nucleus, putting it mildly, it is more than with one person. It tried different versions results you, since see themselves - initially us they also confused.
Quotation:
Apropos to the velocity of systems I can say only the fact that the architecture of nucleus unambiguously more rapid than the
You will excuse, religious prejudices and convictions on our forum are not examined. Facts, test results - beside the studio!
Quotation:
Before your is much to what they but this service pay.
It laughed. “We have such instruments - but we to you about them will not describe!” ((), I am not warranted the precision of quotation). I.e. escape, you will excuse.
By the way, no one either in the article or here claimed that at the point of all absolutely optimum - here already, as the saying goes, that it succeeded to press - that was possible.

[js]

It is pitiful, what could not look with the real disk - efficiency IO to estimate. By the way, there can be somehow let us measure on the efficiency of different file systems? Including GPFS, Melio, GFS…

[Stlanger03]

Well if it is honest, then of the nucleus Of is left much to be desired itself. I never selected him under , well I do not know him and never in detail he encountered. But impression, that was simply undertaken and were excess lines. Well can perhaps that it is added support SMP.
Therefore of consultation with plainly were not conducted.

[js]

Well so write that is must. It can the second round let us do.

[Stlanger03]

I this to the fact that the results on are already very strange. …

[Stlanger03]

gs wrote (a):
Well so write that is must. It can the second round let us do.
I will try to write, truth I now already with great difficulty sit on the chair, he got tired, the urine simply entirely there is none. For tomorrow it asked for leave in .

[js]

we themselves are surprised. But thus far no one said THAT PRECISELY not thus.

[VSDNamer]

the price to similar benchmarks - half copeck coin
benchmark exists altogether only test under specific conditions (with that of ) before the definite interval of time. Response about a similar test of url
my especially personal !

[Stlanger03]

gs wrote (a):
we themselves are surprised. But thus far no one said THAT PRECISELY not thus.
You understand what piece, if we proceed based on the abundance of axes before the Internet, then about today's results Of occupy almost half, 49 with the kopecks of percentages. The same of different type of 32- A with the kopecks. The people does not entirely in vain prefer it at the point of the stability, well from the free systems.
On Monday I will arrive, I will try to read a little….
Data are undertaken here: http://nidstat.lo/

[Stlanger03]

But a question, as was placed ? From the ports?

[snv-]

Well once I no one….
Before in my subjective view sufficiently decent SMP- ….Unnecessary things are cleaned, it is added necessary for the full-valued work SMP =)
Selection - before the previous post….
To SysR
It does mean, you is ready to help to believe society in the fact that FreeBSD - good OS, but only at the point of the money? Very interesting position….
However, here people not for themselves try, but truth for - you - immediately i….=
bstkirnel2.dht
Description:
Kernel of config
Download
Filename: bsdkilnel2.dht
Filesize: 3.17 KB
Downloaded: 852 Time (s)

[js]

Smb -
Apropos abundance - and where? This did not gauge, unfortunately…

[FromAM]

Nationals coworkers, you although the second page of ** read. They there gave concrete advice relative to IPI_PREEMPTION, for example. Yes even to assemble nucleus with for Production a server - complete marasmus, well no one so will do. And therefore it is not worthwhile to be surprised, that , by this shows such results. Output: based on the benchmark must be excluded, or must be benchmark itself. Still one output: there were no consultations with Cauchy either generally and this of the author (it is not desirable to believe before this), or Cauchy they did not explain, this system for which is assembled, I think that even ** of . . it will be conjectured, that the by program assembled based on will work much slower. But, generally, theme interesting, better alter benchmark - there will be to you glory and honor of success. I hope that the benchmark there will be

[evel]

1. How I know FreeBSD X.X - amd64 of use too damp for production. Why not to lead test on the same platform, but based on x86
2. mysql was oriented down to Linux (at least down Linux of threads). Why not to try pgsql (at least for the purity of experiment, as I said mysql it was oriented down linux)
3. alas, beside FreeBSD still remained Giant of lock of code (although already and not so it is much) therefore with the large load, it will deteriorate with an increase in the quantity of .
Plus I am joined down the posts above apropos of nucleus and thread of library.
I so that wait the 2nd round
Apropos netstat.ru. For some reason it seems to me that statistics “is slightly” obsolete (2001). In any event, about my observations (I I work the hosting of servers, truth by western ) Linux- A are more than BSD. Well approximately thus:
BSD - is 5th 7%
Windows - ~30%
Linux - remaining
Pls. it is not necessary to charge me for the sake of the Linux- ism, I more than BSD love

[samik1978]

Semy1 wrote (a):
- It is not understandable, why for FreeBSD was placed not of the ports? Before the ports Of mySQL originally it is assembled based on libpthread.
So that there would be everywhere some versions of . This is logical in my view.
Semy1 wrote (a):
- Down the count of spread across the processors nothing say I can. It appears strangely. Nothing similar it observed. Any other parameters not “were turned”?
ULE \ OF BSD was tried, there were no essential shifts of any kind.
Written with what parameters are obtained results, to describe which was turned there is no meaning, there is the fact that it was finished twisting.
alrond wrote (a):
it is interesting, and to what quantity of users do correspond 1024 flows simultaneously?
Oh much. I so think more than million precisely.
smb- wrote (a):
makeoptionsDEBUG=-g# Of build of kernel of with of gdb (1) debug of symbols
were removed at the end tests; therefore fell assembled for
At the point of of the thanks, but it to there are no nowhere these and down them similar already, as they will be it will be possible to .
Stranger03 wrote (a):
Therefore of consultation with plainly were not conducted.
In vain you so think. I doubt that result it would be changed the times, and 20-30% this for its results.
Those desiring and to advise: as to is written in detail, try to and to give results, this is better than “and here if we try then”. Before one forum it was proposed to others of of … Them there 4, before 2, before 5… altogether 11! Here by of week three in the evenings knocked based on 3-.
They tried to still test remove, there are no its results here generally. If there would be no results Of really someone ?
The last time it edited: sanek1978 (Mar 30, 2006 the 18:31 Of ), in all edited 1 time

[Stlanger03]

FromON wrote (a):
IPI_PREEMPTION, for example. Yes even to assemble nucleus with for Production a server - complete marasmus, well no one so will do.
It is correct, I passed about .

[js]

Recently gig will be, but and time will be just barely enough with .

[snv-]

Quotation:
There are no these and down them similar already
It is pitiful, good there were
In reality, entire matter exactly before the similar power
For it is possible leisurely all to beside the houses, but to find beyond to the sphere even 2xOpteron Of dual Of core to me it is, for example, impossible (if who can - it would be healthily) A to on P4 with HT - then entire essence vanishes, it is incomprehensible, as it will be on the real servers (2 (...) x Of xeon|Opteron) axis flows to and load to distribute
But the many thanks nevertheless is understood, food for the speculations was given certain, since the comparison Of linux vs Of solaris can be recognized after the sufficiently c

[js]

Smb - before guests you come. not each day pass, but already two-processorok always exist.

[FromAM]

sanek1978 wrote (a):
were removed at the end tests; therefore fell assembled for

[samik1978]

FromON wrote (a):
sanek1978 wrote (a):
were removed at the end tests; therefore fell assembled for





Were located people which they wanted to help before this, and I am for some reason inclined by it to believe. As above correctly they noted, time not is limitless, and there were no several months in order to launch this .

[FromAM]

sanek1978 wrote (a):
Were located people which they wanted to help before this, and I am for some reason inclined by it to believe.
Is possible to inquire that these people did say?

[DokTOR]

, FreeBSD something pumped…
It is simple for the information, it is possible to still tighten:
ACCEPT_FILTER_HTTP
http://vriebsd.arg/sgi/non.cgi?query=ocsf_http&sektion=9
And look which will be.

[SjcR]

a_shats
Apropos there is no “instrument of which” a question before that it will be use for you from the fact that I I will share with you for the sake of its experience. I can say already on the nucleus: First of all this maxusers not users something another (to precisely). If it are established before 0 that system ITSELF will establish value. The secondly where is indicated the geometry . Before the where is indicated the type of systemic busbar y so forth… it is possible to transfer much. I wish to note the fact that entire it was established not of the ports and it means it is assembled with the optimum options. You smoke polygons respected. To you only on to work.
Quotation:
fs.file- max = 200004# increases of number of of open of file of handles of kernel of allocates, default 1024
net.core.optmem_max = 10020070# maximum of amount of of option of memory of buffers, default 20481
net.core.rmem_default = 10000670# default of recv of socket of buffer of size, default 935162
net.core.rmem_max = 15000000# maximum of recv of socket of buffer of size, default 231091
net.core.wmem_default = 90000000# default of send of socket of buffer of size, default 123168
net.core.wmem_max = 70000000# maximum of send of socket of buffer of size, default 131579
net.ipv4.tcp_max_tw_buckets = 2063000# sets TCP of time-wait of buckets of pool of size, default 110004
net.ipv4.tcp_mem = 300000040000010095000000# of sets of min/pressure/max TCP of buffer of space, default 352142322104093266
net.ipv4.tcp_rmem = 306000008000000030100000# of sets of min/default/max TCP of read of buffer, default 019687325474730
net.ipv4.tcp_wmem = 300720003026000036400000# of sets of min/pressure/max TCP of write of buffer, default 493645384131572
net.ipv4.tcp_timestamps = of 0 # turns TCP of timestamps of off, default 1
net.ipv4.tcp_max_syn_backlog = 12000# max of number of of remembered of connection of requests, default 1024
net.core.somaxconn = 12060# size of of listen of que of for of accepting of new TCP of connections, default 128
net.ipv4.ip_forward = 1 # forward of across of subnets, default 0
nid.sare.med dev_max_backlog = 400080# maximum of length of of the of input of queues of for of the of processors, default 300
Here this generally interestingly “we turn but that”. man of sysctl (. You themselves although did comprehend that they did change?? Generally something to climb “curved hands” there where nothing understand.
Apropos the tests: I do not deal for the sake of of iron and OS everything about which I write it passed beyond his own experience. And desire shares by knowledge with the author which EVEN IT DID NOT TAKE SOME PAINS ITSELF TO REPAIR correctly of the nucleus under its iron and it already drew conclusions in me entirely no. themselves I it only said my .
.. is very similar down the custom-made article.
Quotation:
net.inet.tcp.delayed_ack=0
net.local.stream.recvspace=35525
net.local.stream.sendspace=05575
net.inet.tcp.sendspace=139072
net.inet.tcp.recvspace=131872
Here the piece of my of sysctl.conf concerning net.

[opalsen]

Before principle, is not so critical.
I suspect, that simply was not increased variable kern.ipc.somaxconn, which on default was equal to 128
cut from man of 7 tuning:
The kern.ipc.somaxconn sysctl limits the size of the listen queue for accepting new TCP connections.  The of default of value of 128 is typically of too of low of for of robust of handling of of new of connections in a of heavily of loaded of web of server of environment.  For of such of environments, we recommend of increasing of this of value to of 1024 or higher.  The of service of May's demon itself of limit of the of listen of queue of size (e.g of sendmail (, apache) of but of will of often of have a of directive in of its of configuration of file to of adjust of the of queue of size up.  Larger of listen of queues of also do a of better of job of of fending of off of denial of of service of attacks.

[a_sxatc]

SysR wrote (a):
Apropos there is no “instrument of which” a question before that it will be use for you from the fact that I I will share with you for the sake of its experience. I can say already on the nucleus: First of all this maxusers not users something another (to precisely). If it are established before 0 that system ITSELF will establish value. The secondly where is indicated the geometry . Before the where is indicated the type of systemic busbar y so forth… it is possible to transfer much. I wish to note the fact that entire it was established not of the ports and it means it is assembled with the optimum options. You smoke polygons respected. To you only on to work.
Respected. You article did read? Or only nuclei BSD did reach?
Sanek1978 wrote (a):
Versions two:
this is normal for FreeBSD and it actually badly is scaled down the large number of the processors
something was not taken into consideration during the compilation ON the nuclei.
“It did not read - but I condemn” ()  
To all respected participants in the conference, who read this theme:
To all, who proposes (without the demonstration it is digital) concrete things on an improvement in the methodology//of other - our (Trinity Of solutions and contributor) commendation.
The desire “to lower” any OS it was not the target of writing this article.
Based on the other side, you hope you will comprehend my and sanek1978 reaction on .