www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - stdio and Tango I/O performance

reply "Andrei Alexandrescu (See Website For Email)" <SeeWebsiteForEmail erdani.org> writes:
Over the past couple of weeks Walter and I have worked on improving the 
stdio module, following my shocking discovery that its performance was 
inadequate for any amount of serious coding. To say nothing about the 
lack of the input routines readf, readln, etc.

The good news is that Walter and myself have improved stdio's reading 
performance by an order of magnitude when compared to a naive 
implementation using fgetc(). Now reading is almost as fast as C would 
allow you, beats Perl and C++, (Perl is harder to beat than C++), while 
writef* is still suboptimal (but write without "f" is fast because it 
forwards to the C routine).

I was actually surprised that nobody seems to have discovered the 
abysmal performance of the writef family. If they did, they would have 
likely posted reports on this newsgroup.

This ate a lot of time measuring, testing, and optimizing. So I have a 
question - has anyone verified that Tango's I/O performance is up to 
snuff? I see it imposes the dynamic-polymorphic approach, and unless 
there was some serious performance work going on, it's possible it's 
even slower than stdio.


Thanks,

Andrei
Mar 21 2007
next sibling parent reply Lars Ivar Igesund <larsivar igesund.net> writes:
Andrei Alexandrescu (See Website For Email) wrote:


 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


If you have actually tested Tango IO and finds it to be slower than
the "new" stdio, then this would be interesting to know, and also the test
harness in use. If not, I find your suggestive tone above to be rather
rude, given the time put into making Tango IO as good as possible. If the
IO in Phobos has been as slow as you say, I highly doubt that Tango has
been in similar ranges, given that it obviates C IO by going directly to
the OS via a buffer.

Now, if you have tested that Tango's IO is slower than Phobos, then maybe
you could spend some time helping to improve it too? Your approach above is
rather unconstructive in an overall fashion.

-- 
Lars Ivar Igesund
blog at http://larsivi.net
DSource, #d.tango & #D: larsivi
Dancing the Tango
Mar 21 2007
parent reply "Andrei Alexandrescu (See Website For Email)" <SeeWebsiteForEmail erdani.org> writes:
Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


 
 If you have actually tested Tango IO and finds it to be slower than
 the "new" stdio, then this would be interesting to know, and also the test
 harness in use. If not, I find your suggestive tone above to be rather
 rude, given the time put into making Tango IO as good as possible. If the
 IO in Phobos has been as slow as you say, I highly doubt that Tango has
 been in similar ranges, given that it obviates C IO by going directly to
 the OS via a buffer.
 
 Now, if you have tested that Tango's IO is slower than Phobos, then maybe
 you could spend some time helping to improve it too? Your approach above is
 rather unconstructive in an overall fashion.


This is a misunderstanding. To date I haven't downloaded Tango.

Andrei
Mar 21 2007
parent reply Lars Ivar Igesund <larsivar igesund.net> writes:
Andrei Alexandrescu (See Website For Email) wrote:


 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


 
 If you have actually tested Tango IO and finds it to be slower than
 the "new" stdio, then this would be interesting to know, and also the
 test harness in use. If not, I find your suggestive tone above to be
 rather rude, given the time put into making Tango IO as good as possible.
 If the IO in Phobos has been as slow as you say, I highly doubt that
 Tango has been in similar ranges, given that it obviates C IO by going
 directly to the OS via a buffer.
 
 Now, if you have tested that Tango's IO is slower than Phobos, then maybe
 you could spend some time helping to improve it too? Your approach above
 is rather unconstructive in an overall fashion.


 
 This is a misunderstanding. To date I haven't downloaded Tango.


Then you should before making any sorts of comments in this vein. Maybe you
even should have compared Tango's IO performance to Phobos' prior to making
such huge efforts to improve on it (Phobos IO)?

-- 
Lars Ivar Igesund
blog at http://larsivi.net
DSource, #d.tango & #D: larsivi
Dancing the Tango
Mar 21 2007
parent reply "Andrei Alexandrescu (See Website For Email)" <SeeWebsiteForEmail erdani.org> writes:
Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:


 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


 If you have actually tested Tango IO and finds it to be slower than
 the "new" stdio, then this would be interesting to know, and also the
 test harness in use. If not, I find your suggestive tone above to be
 rather rude, given the time put into making Tango IO as good as possible.
 If the IO in Phobos has been as slow as you say, I highly doubt that
 Tango has been in similar ranges, given that it obviates C IO by going
 directly to the OS via a buffer.

 Now, if you have tested that Tango's IO is slower than Phobos, then maybe
 you could spend some time helping to improve it too? Your approach above
 is rather unconstructive in an overall fashion.


 This is a misunderstanding. To date I haven't downloaded Tango.


 
 Then you should before making any sorts of comments in this vein. Maybe you
 even should have compared Tango's IO performance to Phobos' prior to making
 such huge efforts to improve on it (Phobos IO)?


Could you please put the gun down. Sheesh. Which comments? What vein? It 
was a question. I simply asked whether there has been performance work 
on Tango's I/O, without having any prior experience with it. I don't see 
reason for anyone to get offended about that.


Andrei
Mar 21 2007
next sibling parent Sean Kelly <sean f4.ca> writes:
Andrei Alexandrescu (See Website For Email) wrote:

 
 Could you please put the gun down. Sheesh. Which comments? What vein? It 
 was a question. I simply asked whether there has been performance work 
 on Tango's I/O, without having any prior experience with it. I don't see 
 reason for anyone to get offended about that.


I think the Tango IO package is all about only paying for what you need. 
  This is one reason for its modular design.  That said, I don't think 
anyone has ever done a performance comparison with C, Phobos, etc.  If 
you're inclined to do so, I would love to see the results.


Sean
Mar 21 2007
prev sibling next sibling parent reply Lars Ivar Igesund <larsivar igesund.net> writes:
Andrei Alexandrescu (See Website For Email) wrote:


 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:


 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


 If you have actually tested Tango IO and finds it to be slower than
 the "new" stdio, then this would be interesting to know, and also the
 test harness in use. If not, I find your suggestive tone above to be
 rather rude, given the time put into making Tango IO as good as
 possible. If the IO in Phobos has been as slow as you say, I highly
 doubt that Tango has been in similar ranges, given that it obviates C
 IO by going directly to the OS via a buffer.

 Now, if you have tested that Tango's IO is slower than Phobos, then
 maybe you could spend some time helping to improve it too? Your
 approach above is rather unconstructive in an overall fashion.


 This is a misunderstanding. To date I haven't downloaded Tango.


 
 Then you should before making any sorts of comments in this vein. Maybe
 you even should have compared Tango's IO performance to Phobos' prior to
 making such huge efforts to improve on it (Phobos IO)?


 
 Could you please put the gun down. Sheesh. Which comments? What vein? It
 was a question. I simply asked whether there has been performance work
 on Tango's I/O, without having any prior experience with it. I don't see
 reason for anyone to get offended about that.


No, I was not offended (not easily done), but I did find your wording
suggestive. And you did give the impression that you had _some_ knowledge
of Tango.

As for that exact question; yes, performance is an important facet of Tango,
IO likewise. Slow IO in Tango would be a defect. And it would be nice to
have an unbiased comparison of the two.

-- 
Lars Ivar Igesund
blog at http://larsivi.net
DSource, #d.tango & #D: larsivi
Dancing the Tango
Mar 21 2007
parent reply "Andrei Alexandrescu (See Website For Email)" <SeeWebsiteForEmail erdani.org> writes:
Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:


 Lars Ivar Igesund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:


 This ate a lot of time measuring, testing, and optimizing. So I have a
 question - has anyone verified that Tango's I/O performance is up to
 snuff? I see it imposes the dynamic-polymorphic approach, and unless
 there was some serious performance work going on, it's possible it's
 even slower than stdio.


 If you have actually tested Tango IO and finds it to be slower than
 the "new" stdio, then this would be interesting to know, and also the
 test harness in use. If not, I find your suggestive tone above to be
 rather rude, given the time put into making Tango IO as good as
 possible. If the IO in Phobos has been as slow as you say, I highly
 doubt that Tango has been in similar ranges, given that it obviates C
 IO by going directly to the OS via a buffer.

 Now, if you have tested that Tango's IO is slower than Phobos, then
 maybe you could spend some time helping to improve it too? Your
 approach above is rather unconstructive in an overall fashion.


 This is a misunderstanding. To date I haven't downloaded Tango.


 Then you should before making any sorts of comments in this vein. Maybe
 you even should have compared Tango's IO performance to Phobos' prior to
 making such huge efforts to improve on it (Phobos IO)?


 Could you please put the gun down. Sheesh. Which comments? What vein? It
 was a question. I simply asked whether there has been performance work
 on Tango's I/O, without having any prior experience with it. I don't see
 reason for anyone to get offended about that.


 
 No, I was not offended (not easily done), but I did find your wording
 suggestive. And you did give the impression that you had _some_ knowledge
 of Tango.
 
 As for that exact question; yes, performance is an important facet of Tango,
 IO likewise. Slow IO in Tango would be a defect. And it would be nice to
 have an unbiased comparison of the two.


Excellent! I have collected some data on stdio and a couple of 
baselines, and I'll install Tango now. In fact I've already hit a hitch: 
  README.txt sends the reader to 
http://dsource.org/projects/tango/wiki/TopicInstallTango which says one 
should run ./lib/install-dmd.sh, which is not to be found in the 
distribution... probably I'll need to set an account on dsource.org and 
ask there.


Andrei
Mar 21 2007
parent reply Sean Kelly <sean f4.ca> writes:
Andrei Alexandrescu (See Website For Email) wrote:

 
 Excellent! I have collected some data on stdio and a couple of 
 baselines, and I'll install Tango now. In fact I've already hit a hitch: 
  README.txt sends the reader to 
 http://dsource.org/projects/tango/wiki/TopicInstallTango which says one 
 should run ./lib/install-dmd.sh, which is not to be found in the 
 distribution... probably I'll need to set an account on dsource.org and 
 ask there.


It was moved to /install/install-dmd.sh, and I guess we missed updating 
the README :-)  If you want to do a manual install, the instructions are 
here: http://www.dsource.org/projects/tango/wiki/TopicInstallTango

Personally, I just publish the appropriate stuff to /include/d/tango and 
/usr/lib directly from the SVN tree via a script.


Sean
Mar 21 2007
parent Sean Kelly <sean f4.ca> writes:
Sean Kelly wrote:

 Andrei Alexandrescu (See Website For Email) wrote:

 Excellent! I have collected some data on stdio and a couple of 
 baselines, and I'll install Tango now. In fact I've already hit a 
 hitch:  README.txt sends the reader to 
 http://dsource.org/projects/tango/wiki/TopicInstallTango which says 
 one should run ./lib/install-dmd.sh, which is not to be found in the 
 distribution... probably I'll need to set an account on dsource.org 
 and ask there.


 
 It was moved to /install/install-dmd.sh, and I guess we missed updating 
 the README :-)  If you want to do a manual install, the instructions are 
 here: http://www.dsource.org/projects/tango/wiki/TopicInstallTango


Oops, I misunderstood.  That link also has instructions for how to 
perform a manual install.
Mar 21 2007
prev sibling parent Walter Bright <newshound digitalmars.com> writes:
Andrei Alexandrescu (See Website For Email) wrote:

 Could you please put the gun down. Sheesh. Which comments? What vein? It 
 was a question. I simply asked whether there has been performance work 
 on Tango's I/O, without having any prior experience with it. I don't see 
 reason for anyone to get offended about that.


Phobos' std.stdio's performance sucked because I'd never done a 
benchmark of it. It was one of those things I cranked out because 
something was needed, intending to go back and optimize later but never 
did...

Anyhow, Andrei's and my experiments with stdio once again showed that 
designing something to be fast and it actually being fast are two very 
different things. We were misled several times by our own preconceived 
notions about what was cheap and what was expensive. Nothing beats 
actually running a benchmark.

The benchmarks we used were all trivial programs, like a loop that just 
reads its stdin, line by line, and then stuffing a 30 Mb file down its 
maw. For example:

import std.stdio;

int main()
{
     char[] buf;
     while (readln(stdin, buf))
      { }
     return 0;
}

Yes, readln() will be in the next update.

Our benchmarking also suggested a couple performance related 
improvements to the gc interface, which I am working on.
Mar 21 2007
prev sibling parent reply =?ISO-8859-1?Q?Anders_F_Bj=F6rklund?= <afb algonet.se> writes:
Andrei Alexandrescu (See Website For Email) wrote:


 Over the past couple of weeks Walter and I have worked on improving the 
 stdio module, following my shocking discovery that its performance was 
 inadequate for any amount of serious coding. To say nothing about the 
 lack of the input routines readf, readln, etc.


It's excellent news that std.stdio is finally getting revamped!
Hopefully you were able to make some use of our previous code
and suggestions, from the discussions held earlier (in 2005).


 I was actually surprised that nobody seems to have discovered the 
 abysmal performance of the writef family. If they did, they would have 
 likely posted reports on this newsgroup.


I'm afraid that at least some were more worried about the missing
features initially, and figured that performance would come "later".
Or even that one should use the std.c.stdio functions instead...

For the greatest of times, printf was the recommended output method.
(for instance, looking at "hello.d" in DMD 1.009 shows printf use.
Looking at the "pi.d" example, and it shows using using sscanf ?)

BTW: I don't think there is anything wrong with using "C" I/O routines,
on the contrary it is very useful when quickly porting programs over,
but it does suggest that "D" I/O methods in Phobos were not up to it ?

--anders
Mar 21 2007
parent "Andrei Alexandrescu (See Website For Email)" <SeeWebsiteForEmail erdani.org> writes:
Anders F Björklund wrote:

 Andrei Alexandrescu (See Website For Email) wrote:
 

 Over the past couple of weeks Walter and I have worked on improving 
 the stdio module, following my shocking discovery that its performance 
 was inadequate for any amount of serious coding. To say nothing about 
 the lack of the input routines readf, readln, etc.


 
 It's excellent news that std.stdio is finally getting revamped!
 Hopefully you were able to make some use of our previous code
 and suggestions, from the discussions held earlier (in 2005).


The only concern is the risk of concurrent development of stdio and 
Tango's I/O. It would be best to not balkanize the community by 
presenting them functionally almost equivalent interfaces. Phobos 
already has two, no, three mutually incompatible ways of dealing with 
files. I don't know how this can be approached effectively.


Andrei
Mar 21 2007