Re: [9fans] FTQ benchmark available

Re: [9fans] FTQ benchmark available

[Matthew Sottile]

> I don't know; I didn't know of the existence of any such thing.
> Ron just handed me a tarball one day and said, "Thou Shalt Port This",
> so I did. I'll look at the one on sourceforge.
>
> John

Yep - a basic plan9 version was added to the repo a year or two ago.
I haven't tested it at all though, and I haven't made sure that
changes to the core of FTQ made it into the plan9-specific sections.
You might take a look at it.

-m

Re: [9fans] FTQ benchmark available

[dave.l]

> Actually, in Real Computing, aka "not on your desktop", some groups
> such as, oh, IBM and the US Government (you may have heard of them)
> care about how fast a OS can perform in this kind of thing. They don't
> care about word processors, spreadsheets, web browsers, any of the
> stuff Joe Windowsuser will bitch about; what they really want is fast
> execution of parallel programs... and probably FORTRAN 77.

You and Ron are right.

Also, I didn't really understand what FTQ was trying to measure ...

I'll shut up and go write some drivers.

DaveL



I'm jumping off the top of Guy's Hospital
to support the Myasthenia Gravis Association.

Please support me!

http://www.justgiving.com/davelukes-abseil

Re: [9fans] FTQ benchmark available

[Charles Forsyth]

> Seriously: what's the point?

i'm sometimes interested in seeing the differences (if any) caused
by certain system changes, with a view to showing that you can be
fast but still straightforward and sane, without a load of ugly code
let alone inlined assembly code and 60,000 lines of include files.

Re: [9fans] FTQ benchmark available

[John Floren]

On Fri, Apr 11, 2008 at 5:23 PM,  <dave.l@mac.com> wrote:
>
> > measurement is hard. Let's go write drivers -- Malibu Barbie.
> >
>
>  Good Idea.
>
>  I don't give a flying f**k how slow it is: it's still a million times
> better than (random other OS) -- Malibu Dave.
>
>  Seriously: what's the point?
>
>  Even if you "prove" that plan9 is "faster" than brandX,
>  it still won't have a word processor, spreadsheet, DBMS, shared objects,
> eyecandy, jessica_biel_naked_in_my_bed.c
>  and whatever else is  "vital" if an OS is to "succeed" in the "real world".
>

Actually, in Real Computing, aka "not on your desktop", some groups
such as, oh, IBM and the US Government (you may have heard of them)
care about how fast a OS can perform in this kind of thing. They don't
care about word processors, spreadsheets, web browsers, any of the
stuff Joe Windowsuser will bitch about; what they really want is fast
execution of parallel programs... and probably FORTRAN 77.



John
--
Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn

Re: [9fans] FTQ benchmark available

[ron minnich]

On Fri, Apr 11, 2008 at 5:23 PM,  <dave.l@mac.com> wrote:

>  Even if you "prove" that plan9 is "faster" than brandX,

The point is that I use this data in my work. Whether you wish to use
it is up to you.

Some folks have actually made use of this code, e.g. Cray used it to
tune the OS that runs on the 20K CPU XT4.

Yeah, the code seems simple and stupid, but there are a few subtleties
in there, and it may surprise you (it sure surprised me) that this
code is the first code I know of that got those subtleties right. It
was a lot of work and most of it fell on The Math Dude (not me!) to
make sure it was right.

now go write those drivers, and watch yourself if Ken shows up.

ron

Re: [9fans] FTQ benchmark available

[dave.l]

> measurement is hard. Let's go write drivers -- Malibu Barbie.

Good Idea.

I don't give a flying f**k how slow it is: it's still a million times
better than (random other OS) -- Malibu Dave.

Seriously: what's the point?

Even if you "prove" that plan9 is "faster" than brandX,
it still won't have a word processor, spreadsheet, DBMS, shared
objects, eyecandy, jessica_biel_naked_in_my_bed.c
and whatever else is  "vital" if an OS is to "succeed" in the "real
world".

DaveL

I'm jumping off the top of Guy's Hospital
to support the Myasthenia Gravis Association.

Please support me!

http://www.justgiving.com/davelukes-abseil

Re: [9fans] FTQ benchmark available

[ron minnich]

On Thu, Apr 10, 2008 at 7:15 PM, erik quanstrom <quanstro@quanstro.net> wrote:
> > The two upticks on the top one usually mean trouble with the timer. I
>  > can't imagine why it is happening.
>
>  missyncronized tsc?
>

Good possibility, but without knowing what it was run on it's hard to
say. Obviously, we have some work to do on plan 9. We should really
check out #p, then fire up enough procs and wire them down.

ron

Re: [9fans] FTQ benchmark available

[erik quanstrom]

> The two upticks on the top one usually mean trouble with the timer. I
> can't imagine why it is happening.

missyncronized tsc?

- erik

Re: [9fans] FTQ benchmark available

[ron minnich]

On Thu, Apr 10, 2008 at 2:05 PM, andrey mirtchovski
<mirtchovski@gmail.com> wrote:
> >  i wonder if they'll give the same results!
>
>  yes, the results are very similar (it's not an exact number). top here
>  is the original ftq, bottom is John's port:
>
>  http://mirtchovski.com/screenshots/ftq.jpg
>
>  btw, John, please add a "clean" option to the mkfile; also,
>  exits("success") is wrong.
>
>

The two upticks on the top one usually mean trouble with the timer. I
can't imagine why it is happening.

measurement is hard. Let's go write drivers -- Malibu Barbie.

ron

Re: [9fans] FTQ benchmark available

[andrey mirtchovski]

the only major difference would be between the respective
"getticks()" and "cycles()" implementations, i think.

On Thu, Apr 10, 2008 at 3:28 PM,  <john@csplan9.rit.edu> wrote:
>
> >>  i wonder if they'll give the same results!
>  >
>  > yes, the results are very similar (it's not an exact number). top here
>  > is the original ftq, bottom is John's port:
>  >
>  > http://mirtchovski.com/screenshots/ftq.jpg
>  >
>  > btw, John, please add a "clean" option to the mkfile; also,
>  > exits("success") is wrong.
>
>  Ok, fixed that stuff and ran some tests of my own.
>  My ftq: http://csplan9.rit.edu/users/john/my-ftq.png
>  Original ftq: http://csplan9.rit.edu/users/john/other-ftq.png
>
>  The FFTs look very different; I wonder which one is more correct?
>  I ran these tests one after the other on our CPU/auth/file server,
>  which had only one user (me) logged in over drawterm.
>
>  John
>
>
>

Re: [9fans] FTQ benchmark available

[john]

>>  i wonder if they'll give the same results!
>
> yes, the results are very similar (it's not an exact number). top here
> is the original ftq, bottom is John's port:
>
> http://mirtchovski.com/screenshots/ftq.jpg
>
> btw, John, please add a "clean" option to the mkfile; also,
> exits("success") is wrong.

Ok, fixed that stuff and ran some tests of my own.
My ftq: http://csplan9.rit.edu/users/john/my-ftq.png
Original ftq: http://csplan9.rit.edu/users/john/other-ftq.png

The FFTs look very different; I wonder which one is more correct?
I ran these tests one after the other on our CPU/auth/file server,
which had only one user (me) logged in over drawterm.

John

Re: [9fans] FTQ benchmark available

[andrey mirtchovski]

>  i wonder if they'll give the same results!

yes, the results are very similar (it's not an exact number). top here
is the original ftq, bottom is John's port:

http://mirtchovski.com/screenshots/ftq.jpg

btw, John, please add a "clean" option to the mkfile; also,
exits("success") is wrong.

Re: [9fans] FTQ benchmark available

[ron minnich]

On Thu, Apr 10, 2008 at 11:25 AM, Charles Forsyth <forsyth@terzarima.net> wrote:
> > Ron just handed me a tarball one day and said, "Thou Shalt Port This",
>  > so I did. I'll look at the one on sourceforge.
>
>  i wonder if they'll give the same results!
>
>

actually, that's a good test. The worst part is that andrey has told
me many times that the port was out there. I claim advanced brain cell
death on my part

ron

Re: [9fans] FTQ benchmark available

[Charles Forsyth]

> Ron just handed me a tarball one day and said, "Thou Shalt Port This",
> so I did. I'll look at the one on sourceforge.

i wonder if they'll give the same results!

Re: [9fans] FTQ benchmark available

[ron minnich]

On Thu, Apr 10, 2008 at 10:28 AM,  <john@csplan9.rit.edu> wrote:
> > how is that different from the Plan 9 port that is already provided
>  > with the ftq (chamatools) distribution from sourceforge?
>  >
>
>  I don't know; I didn't know of the existence of any such thing.
>  Ron just handed me a tarball one day and said, "Thou Shalt Port This",
>  so I did. I'll look at the one on sourceforge.
>

yes, well, ron is forgetful sometimes.

what can I say?

ron

Re: [9fans] FTQ benchmark available

[john]

> how is that different from the Plan 9 port that is already provided
> with the ftq (chamatools) distribution from sourceforge?
>

I don't know; I didn't know of the existence of any such thing.
Ron just handed me a tarball one day and said, "Thou Shalt Port This",
so I did. I'll look at the one on sourceforge.

John

Re: [9fans] FTQ benchmark available

[andrey mirtchovski]

how is that different from the Plan 9 port that is already provided
with the ftq (chamatools) distribution from sourceforge?

On Thu, Apr 10, 2008 at 11:09 AM,  <john@csplan9.rit.edu> wrote:
> The Finite Time Quantum benchmark has been ported to Plan 9; just mk
>  it and go.  There is a python program included to analyze and plot the
>  data which you should probably run on Linux.  FTQ measures operating
>  system "noise" by trying to do as much work as possible in fixed time
>  sections; if the operating system is "noisy", less work will be done
>  and you'll see spikes on the graph.  The best way to understand this
>  is to run the test and look at it.
>
>  Tarball is at /n/sources/contrib/john/ftq.tgz
>
>  Unlike the lmbench port, there are no restrictions on this
>  benchmark--go ahead and share your results, compare them to Linux,
>  etc.
>
>
>  John
>
>
>

[9fans] FTQ benchmark available

[john]

The Finite Time Quantum benchmark has been ported to Plan 9; just mk
it and go.  There is a python program included to analyze and plot the
data which you should probably run on Linux.  FTQ measures operating
system "noise" by trying to do as much work as possible in fixed time
sections; if the operating system is "noisy", less work will be done
and you'll see spikes on the graph.  The best way to understand this
is to run the test and look at it.

Tarball is at /n/sources/contrib/john/ftq.tgz

Unlike the lmbench port, there are no restrictions on this
benchmark--go ahead and share your results, compare them to Linux,
etc.


John