Michael Snoyman | 8 Mar 07:30 2013

File I/O benchmark help (conduit, io-streams and Handle)

Hi all,

I'm turning to the community for some help understanding some benchmark results[1]. I was curious to see how the new io-streams would work with conduit, as it looks like a far saner low-level approach than Handles. In fact, the API is so simple that the entire wrapper is just a few lines of code[2].


I then added in some basic file copy benchmarks, comparing conduit+Handle (with ResourceT or bracket), conduit+io-streams, straight io-streams, and lazy I/O. All approaches fell into the same ballpark, with conduit+bracket and conduit+io-streams taking a slight lead. (I haven't analyzed that enough to know if it means anything, however.)

Then I decided to pull up the NoHandle code I wrote a while ago for conduit. This code was written initially for Windows only, to work around the fact that System.IO.openFile does some file locking. To avoid using Handles, I wrote a simple FFI wrapper exposing open, read, and close system calls, ported it to POSIX, and hid it behind a Cabal flag. Out of curiosity, I decided to expose it and include it in the benchmark.

The results are extreme. I've confirmed multiple times that the copy algorithm is in fact copying the file, so I don't think the test itself is cheating somehow. But I don't know how to explain the massive gap. I've run this on two different systems. The results you see linked are from my local machine. On an EC2 instance, the gap was a bit smaller, but the NoHandle code was still 75% faster than the others.

My initial guess is that I'm not properly tying into the IO manager, but I wanted to see if the community had any thoughts. The relevant pieces of code are [3][4][5].

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Michael Snoyman | 8 Mar 07:50 2013

Re: File I/O benchmark help (conduit, io-streams and Handle)

One clarification: it seems that sourceFile and sourceFileNoHandle have virtually no difference in speed. The gap comes exclusively from sinkFile vs sinkFileNoHandle. This makes me think that it might be a buffer copy that's causing the slowdown, in which case the benchmark may in fact be accurate.

On Mar 8, 2013 8:30 AM, "Michael Snoyman" <michael <at> snoyman.com> wrote:
Hi all,

I'm turning to the community for some help understanding some benchmark results[1]. I was curious to see how the new io-streams would work with conduit, as it looks like a far saner low-level approach than Handles. In fact, the API is so simple that the entire wrapper is just a few lines of code[2].


I then added in some basic file copy benchmarks, comparing conduit+Handle (with ResourceT or bracket), conduit+io-streams, straight io-streams, and lazy I/O. All approaches fell into the same ballpark, with conduit+bracket and conduit+io-streams taking a slight lead. (I haven't analyzed that enough to know if it means anything, however.)

Then I decided to pull up the NoHandle code I wrote a while ago for conduit. This code was written initially for Windows only, to work around the fact that System.IO.openFile does some file locking. To avoid using Handles, I wrote a simple FFI wrapper exposing open, read, and close system calls, ported it to POSIX, and hid it behind a Cabal flag. Out of curiosity, I decided to expose it and include it in the benchmark.

The results are extreme. I've confirmed multiple times that the copy algorithm is in fact copying the file, so I don't think the test itself is cheating somehow. But I don't know how to explain the massive gap. I've run this on two different systems. The results you see linked are from my local machine. On an EC2 instance, the gap was a bit smaller, but the NoHandle code was still 75% faster than the others.

My initial guess is that I'm not properly tying into the IO manager, but I wanted to see if the community had any thoughts. The relevant pieces of code are [3][4][5].

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
John Lato | 8 Mar 07:59 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

I would have expected sourceFileNoHandle to make the most difference, since that's one location (write) where you've obviously removed a copy. Does sourceFileNoHandle allocate less?

Incidentally, I've recently been making similar changes to IO code (removing buffer copies) and getting similar speedups.  Although the results tend to be less pronounced in code that isn't strictly IO-bound.


On Fri, Mar 8, 2013 at 2:50 PM, Michael Snoyman <michael <at> snoyman.com> wrote:

One clarification: it seems that sourceFile and sourceFileNoHandle have virtually no difference in speed. The gap comes exclusively from sinkFile vs sinkFileNoHandle. This makes me think that it might be a buffer copy that's causing the slowdown, in which case the benchmark may in fact be accurate.

On Mar 8, 2013 8:30 AM, "Michael Snoyman" <michael <at> snoyman.com> wrote:
Hi all,

I'm turning to the community for some help understanding some benchmark results[1]. I was curious to see how the new io-streams would work with conduit, as it looks like a far saner low-level approach than Handles. In fact, the API is so simple that the entire wrapper is just a few lines of code[2].


I then added in some basic file copy benchmarks, comparing conduit+Handle (with ResourceT or bracket), conduit+io-streams, straight io-streams, and lazy I/O. All approaches fell into the same ballpark, with conduit+bracket and conduit+io-streams taking a slight lead. (I haven't analyzed that enough to know if it means anything, however.)

Then I decided to pull up the NoHandle code I wrote a while ago for conduit. This code was written initially for Windows only, to work around the fact that System.IO.openFile does some file locking. To avoid using Handles, I wrote a simple FFI wrapper exposing open, read, and close system calls, ported it to POSIX, and hid it behind a Cabal flag. Out of curiosity, I decided to expose it and include it in the benchmark.

The results are extreme. I've confirmed multiple times that the copy algorithm is in fact copying the file, so I don't think the test itself is cheating somehow. But I don't know how to explain the massive gap. I've run this on two different systems. The results you see linked are from my local machine. On an EC2 instance, the gap was a bit smaller, but the NoHandle code was still 75% faster than the others.

My initial guess is that I'm not properly tying into the IO manager, but I wanted to see if the community had any thoughts. The relevant pieces of code are [3][4][5].


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Gregory Collins | 8 Mar 09:36 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

+Simon Marlow
A couple of comments:
  • maybe we shouldn't back the file by a Handle. io-streams does this by default out of the box; I had a posix file interface for unix (guarded by CPP) for a while but decided to ditch it for simplicity. If your results are correct, given how slow going by Handle seems to be I may revisit this, I figured it would be "good enough".
  • io-streams turns Handle buffering off in withFileAsOutput. So the difference shouldn't be as a result of buffering. Simon: is this an expected result? I presume you did some Handle debugging?
  • the IO manager should not have any bearing here because file code doesn't actually ever use it (epoll() doesn't work for files)
  • does the difference persist when the file size gets bigger?
  • your file descriptor code doesn't handle EINTR properly, although you said you checked that the file copy is being done?
  • Copying a 1MB file in 1ms gives a throughput of ~1GB/s. The other methods have a more believable ~70MB/s throughput.
G


On Fri, Mar 8, 2013 at 7:30 AM, Michael Snoyman <michael <at> snoyman.com> wrote:
Hi all,

I'm turning to the community for some help understanding some benchmark results[1]. I was curious to see how the new io-streams would work with conduit, as it looks like a far saner low-level approach than Handles. In fact, the API is so simple that the entire wrapper is just a few lines of code[2].


I then added in some basic file copy benchmarks, comparing conduit+Handle (with ResourceT or bracket), conduit+io-streams, straight io-streams, and lazy I/O. All approaches fell into the same ballpark, with conduit+bracket and conduit+io-streams taking a slight lead. (I haven't analyzed that enough to know if it means anything, however.)

Then I decided to pull up the NoHandle code I wrote a while ago for conduit. This code was written initially for Windows only, to work around the fact that System.IO.openFile does some file locking. To avoid using Handles, I wrote a simple FFI wrapper exposing open, read, and close system calls, ported it to POSIX, and hid it behind a Cabal flag. Out of curiosity, I decided to expose it and include it in the benchmark.

The results are extreme. I've confirmed multiple times that the copy algorithm is in fact copying the file, so I don't think the test itself is cheating somehow. But I don't know how to explain the massive gap. I've run this on two different systems. The results you see linked are from my local machine. On an EC2 instance, the gap was a bit smaller, but the NoHandle code was still 75% faster than the others.

My initial guess is that I'm not properly tying into the IO manager, but I wanted to see if the community had any thoughts. The relevant pieces of code are [3][4][5].


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe




--
Gregory Collins <greg <at> gregorycollins.net>
_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Gregory Collins | 8 Mar 09:37 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)


On Fri, Mar 8, 2013 at 9:36 AM, Gregory Collins <greg <at> gregorycollins.net> wrote:
I presume you did some Handle debugging?

...and here I mean "benchmarking" of course.


--
Gregory Collins <greg <at> gregorycollins.net>
_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
John Lato | 8 Mar 09:48 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

I'd like to point out that it's entirely possible to get good performance out of a handle.     The iteratee package has had both FD and Handle-based IO for a while, and I've never observed any serious performance differences between the two.  Also, if I may be so bold, Michael's supercharged copy speeds are on par with iteratee's performance using Handles:

So while there's definitely something interesting going on here, I think it needs a bit more investigation before suggesting that Handles should be avoided.

For comparison, on my system I get 
$ time cp input.dat output.dat 

real 0m0.004s
user 0m0.000s
sys 0m0.000s

so the throughput observed on the faster times is entirely reasonable.

John L.


On Fri, Mar 8, 2013 at 4:36 PM, Gregory Collins <greg <at> gregorycollins.net> wrote:
+Simon Marlow
A couple of comments:
  • maybe we shouldn't back the file by a Handle. io-streams does this by default out of the box; I had a posix file interface for unix (guarded by CPP) for a while but decided to ditch it for simplicity. If your results are correct, given how slow going by Handle seems to be I may revisit this, I figured it would be "good enough".
  • io-streams turns Handle buffering off in withFileAsOutput. So the difference shouldn't be as a result of buffering. Simon: is this an expected result? I presume you did some Handle debugging?
  • the IO manager should not have any bearing here because file code doesn't actually ever use it (epoll() doesn't work for files)
  • does the difference persist when the file size gets bigger?
  • your file descriptor code doesn't handle EINTR properly, although you said you checked that the file copy is being done?
  • Copying a 1MB file in 1ms gives a throughput of ~1GB/s. The other methods have a more believable ~70MB/s throughput.
G


On Fri, Mar 8, 2013 at 7:30 AM, Michael Snoyman <michael <at> snoyman.com> wrote:
Hi all,

I'm turning to the community for some help understanding some benchmark results[1]. I was curious to see how the new io-streams would work with conduit, as it looks like a far saner low-level approach than Handles. In fact, the API is so simple that the entire wrapper is just a few lines of code[2].


I then added in some basic file copy benchmarks, comparing conduit+Handle (with ResourceT or bracket), conduit+io-streams, straight io-streams, and lazy I/O. All approaches fell into the same ballpark, with conduit+bracket and conduit+io-streams taking a slight lead. (I haven't analyzed that enough to know if it means anything, however.)

Then I decided to pull up the NoHandle code I wrote a while ago for conduit. This code was written initially for Windows only, to work around the fact that System.IO.openFile does some file locking. To avoid using Handles, I wrote a simple FFI wrapper exposing open, read, and close system calls, ported it to POSIX, and hid it behind a Cabal flag. Out of curiosity, I decided to expose it and include it in the benchmark.

The results are extreme. I've confirmed multiple times that the copy algorithm is in fact copying the file, so I don't think the test itself is cheating somehow. But I don't know how to explain the massive gap. I've run this on two different systems. The results you see linked are from my local machine. On an EC2 instance, the gap was a bit smaller, but the NoHandle code was still 75% faster than the others.

My initial guess is that I'm not properly tying into the IO manager, but I wanted to see if the community had any thoughts. The relevant pieces of code are [3][4][5].


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe




--
Gregory Collins <greg <at> gregorycollins.net>

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Gregory Collins | 8 Mar 09:53 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

On Fri, Mar 8, 2013 at 9:48 AM, John Lato <jwlato <at> gmail.com> wrote:
For comparison, on my system I get 
$ time cp input.dat output.dat 

real 0m0.004s
user 0m0.000s
sys 0m0.000s

Does your workstation have an SSD? Michael's using a spinning disk.


--
Gregory Collins <greg <at> gregorycollins.net>
_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Alexander Kjeldaas | 8 Mar 10:13 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)




On Fri, Mar 8, 2013 at 9:53 AM, Gregory Collins <greg <at> gregorycollins.net> wrote:
On Fri, Mar 8, 2013 at 9:48 AM, John Lato <jwlato <at> gmail.com> wrote:
For comparison, on my system I get 
$ time cp input.dat output.dat 

real 0m0.004s
user 0m0.000s
sys 0m0.000s

Does your workstation have an SSD? Michael's using a spinning disk.


If you're only copying a GB or so, it should only be memory traffic.

Alexander
 

--
Gregory Collins <greg <at> gregorycollins.net>

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Gregory Collins | 8 Mar 11:13 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

Something must be wrong with the conduit "NoHandle" code. I increased the filesize to 60MB and implemented the copy loop in pure C, the code and results are here:


Everything but the conduit NoHandle code runs in roughly 600-620ms, including the pure C version.

G


On Fri, Mar 8, 2013 at 10:13 AM, Alexander Kjeldaas <alexander.kjeldaas <at> gmail.com> wrote:



On Fri, Mar 8, 2013 at 9:53 AM, Gregory Collins <greg <at> gregorycollins.net> wrote:
On Fri, Mar 8, 2013 at 9:48 AM, John Lato <jwlato <at> gmail.com> wrote:
For comparison, on my system I get 
$ time cp input.dat output.dat 

real 0m0.004s
user 0m0.000s
sys 0m0.000s

Does your workstation have an SSD? Michael's using a spinning disk.


If you're only copying a GB or so, it should only be memory traffic.

Alexander
 

--
Gregory Collins <greg <at> gregorycollins.net>

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe





--
Gregory Collins <greg <at> gregorycollins.net>
_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Michael Snoyman | 8 Mar 11:32 2013

Re: File I/O benchmark help (conduit, io-streams and Handle)

That demonstrated the issue: I'd forgotten to pass O_TRUNC to the open system call. Adding that back makes the numbers much more comparable.

Thanks for the input everyone, and Gregory for finding the actual problem (as well as pointing out a few other improvements).


On Fri, Mar 8, 2013 at 12:13 PM, Gregory Collins <greg <at> gregorycollins.net> wrote:
Something must be wrong with the conduit "NoHandle" code. I increased the filesize to 60MB and implemented the copy loop in pure C, the code and results are here:


Everything but the conduit NoHandle code runs in roughly 600-620ms, including the pure C version.

G


On Fri, Mar 8, 2013 at 10:13 AM, Alexander Kjeldaas <alexander.kjeldaas <at> gmail.com> wrote:



On Fri, Mar 8, 2013 at 9:53 AM, Gregory Collins <greg <at> gregorycollins.net> wrote:
On Fri, Mar 8, 2013 at 9:48 AM, John Lato <jwlato <at> gmail.com> wrote:
For comparison, on my system I get 
$ time cp input.dat output.dat 

real 0m0.004s
user 0m0.000s
sys 0m0.000s

Does your workstation have an SSD? Michael's using a spinning disk.


If you're only copying a GB or so, it should only be memory traffic.

Alexander
 

--
Gregory Collins <greg <at> gregorycollins.net>

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe





--
Gregory Collins <greg <at> gregorycollins.net>

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Simon Marlow | 8 Mar 11:36 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

1GB/s for copying a file is reasonable - it's around half the memory 
bandwidth, so copying the data twice would give that result (assuming no 
actual I/O is taking place, which is what you want because actual I/O 
will swamp any differences at the software level).

The Handle overhead should be negligible if you're only using 
hGetBufSome and hPutBuf, because those functions basically just call 
read() and write() when the amount of data is larger than the buffer size.

There's clearly something suspicious going on here, unfortunately I 
don't have time right now to investigate, but I'll keep an eye on the 
thread.

Cheers,
	Simon

On 08/03/13 08:36, Gregory Collins wrote:
> +Simon Marlow
> A couple of comments:
>
>   * maybe we shouldn't back the file by a Handle. io-streams does this
>     by default out of the box; I had a posix file interface for unix
>     (guarded by CPP) for a while but decided to ditch it for simplicity.
>     If your results are correct, given how slow going by Handle seems to
>     be I may revisit this, I figured it would be "good enough".
>   * io-streams turns Handle buffering off in withFileAsOutput. So the
>     difference shouldn't be as a result of buffering. Simon: is this an
>     expected result? I presume you did some Handle debugging?
>   * the IO manager should not have any bearing here because file code
>     doesn't actually ever use it (epoll() doesn't work for files)
>   * does the difference persist when the file size gets bigger?
>   * your file descriptor code doesn't handle EINTR properly, although
>     you said you checked that the file copy is being done?
>   * Copying a 1MB file in 1ms gives a throughput of ~1GB/s. The other
>     methods have a more believable ~70MB/s throughput.
>
> G
>
>
> On Fri, Mar 8, 2013 at 7:30 AM, Michael Snoyman <michael <at> snoyman.com
> <mailto:michael <at> snoyman.com>> wrote:
>
>     Hi all,
>
>     I'm turning to the community for some help understanding some
>     benchmark results[1]. I was curious to see how the new io-streams
>     would work with conduit, as it looks like a far saner low-level
>     approach than Handles. In fact, the API is so simple that the entire
>     wrapper is just a few lines of code[2].
>
>     I then added in some basic file copy benchmarks, comparing
>     conduit+Handle (with ResourceT or bracket), conduit+io-streams,
>     straight io-streams, and lazy I/O. All approaches fell into the same
>     ballpark, with conduit+bracket and conduit+io-streams taking a
>     slight lead. (I haven't analyzed that enough to know if it means
>     anything, however.)
>
>     Then I decided to pull up the NoHandle code I wrote a while ago for
>     conduit. This code was written initially for Windows only, to work
>     around the fact that System.IO.openFile does some file locking. To
>     avoid using Handles, I wrote a simple FFI wrapper exposing open,
>     read, and close system calls, ported it to POSIX, and hid it behind
>     a Cabal flag. Out of curiosity, I decided to expose it and include
>     it in the benchmark.
>
>     The results are extreme. I've confirmed multiple times that the copy
>     algorithm is in fact copying the file, so I don't think the test
>     itself is cheating somehow. But I don't know how to explain the
>     massive gap. I've run this on two different systems. The results you
>     see linked are from my local machine. On an EC2 instance, the gap
>     was a bit smaller, but the NoHandle code was still 75% faster than
>     the others.
>
>     My initial guess is that I'm not properly tying into the IO manager,
>     but I wanted to see if the community had any thoughts. The relevant
>     pieces of code are [3][4][5].
>
>     Michael
>
>     [1] http://static.snoyman.com/streams.html
>     [2]
>     https://github.com/snoyberg/conduit/blob/streams/io-streams-conduit/Data/Conduit/Streams.hs
>     [3]
>     https://github.com/snoyberg/conduit/blob/streams/conduit/System/PosixFile.hsc
>     [4]
>     https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L54
>     [5]
>     https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L167
>
>     _______________________________________________
>     Haskell-Cafe mailing list
>     Haskell-Cafe <at> haskell.org <mailto:Haskell-Cafe <at> haskell.org>
>     http://www.haskell.org/mailman/listinfo/haskell-cafe
>
>
>
>
> --
> Gregory Collins <greg <at> gregorycollins.net <mailto:greg <at> gregorycollins.net>>
John Lato | 9 Mar 09:01 2013
Picon

Re: File I/O benchmark help (conduit, io-streams and Handle)

On Fri, Mar 8, 2013 at 6:36 PM, Simon Marlow <marlowsd <at> gmail.com> wrote:
1GB/s for copying a file is reasonable - it's around half the memory bandwidth, so copying the data twice would give that result (assuming no actual I/O is taking place, which is what you want because actual I/O will swamp any differences at the software level).

The Handle overhead should be negligible if you're only using hGetBufSome and hPutBuf, because those functions basically just call read() and write() when the amount of data is larger than the buffer size.

There's clearly something suspicious going on here, unfortunately I don't have time right now to investigate, but I'll keep an eye on the thread.

Possibly disk caching/syncing issues?  If some of the tests are able to either read entirely from cache (on the 1MB test), or don't completely sync after the write, they could happen much faster than others that have to actually hit the disk.  For the 60MB test, it's almost guaranteed that actual IO would take place and dominate the timings.

John L.


Cheers,
        Simon


On 08/03/13 08:36, Gregory Collins wrote:
+Simon Marlow
A couple of comments:

  * maybe we shouldn't back the file by a Handle. io-streams does this

    by default out of the box; I had a posix file interface for unix
    (guarded by CPP) for a while but decided to ditch it for simplicity.
    If your results are correct, given how slow going by Handle seems to
    be I may revisit this, I figured it would be "good enough".
  * io-streams turns Handle buffering off in withFileAsOutput. So the

    difference shouldn't be as a result of buffering. Simon: is this an
    expected result? I presume you did some Handle debugging?
  * the IO manager should not have any bearing here because file code

    doesn't actually ever use it (epoll() doesn't work for files)
  * does the difference persist when the file size gets bigger?
  * your file descriptor code doesn't handle EINTR properly, although

    you said you checked that the file copy is being done?
  * Copying a 1MB file in 1ms gives a throughput of ~1GB/s. The other

    methods have a more believable ~70MB/s throughput.

G


On Fri, Mar 8, 2013 at 7:30 AM, Michael Snoyman <michael <at> snoyman.com
<mailto:michael <at> snoyman.com>> wrote:

    Hi all,

    I'm turning to the community for some help understanding some
    benchmark results[1]. I was curious to see how the new io-streams
    would work with conduit, as it looks like a far saner low-level
    approach than Handles. In fact, the API is so simple that the entire
    wrapper is just a few lines of code[2].

    I then added in some basic file copy benchmarks, comparing
    conduit+Handle (with ResourceT or bracket), conduit+io-streams,
    straight io-streams, and lazy I/O. All approaches fell into the same
    ballpark, with conduit+bracket and conduit+io-streams taking a
    slight lead. (I haven't analyzed that enough to know if it means
    anything, however.)

    Then I decided to pull up the NoHandle code I wrote a while ago for
    conduit. This code was written initially for Windows only, to work
    around the fact that System.IO.openFile does some file locking. To
    avoid using Handles, I wrote a simple FFI wrapper exposing open,
    read, and close system calls, ported it to POSIX, and hid it behind
    a Cabal flag. Out of curiosity, I decided to expose it and include
    it in the benchmark.

    The results are extreme. I've confirmed multiple times that the copy
    algorithm is in fact copying the file, so I don't think the test
    itself is cheating somehow. But I don't know how to explain the
    massive gap. I've run this on two different systems. The results you
    see linked are from my local machine. On an EC2 instance, the gap
    was a bit smaller, but the NoHandle code was still 75% faster than
    the others.

    My initial guess is that I'm not properly tying into the IO manager,
    but I wanted to see if the community had any thoughts. The relevant
    pieces of code are [3][4][5].

    Michael

    [1] http://static.snoyman.com/streams.html
    [2]
    https://github.com/snoyberg/conduit/blob/streams/io-streams-conduit/Data/Conduit/Streams.hs
    [3]
    https://github.com/snoyberg/conduit/blob/streams/conduit/System/PosixFile.hsc
    [4]
    https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L54
    [5]
    https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L167

    _______________________________________________
    Haskell-Cafe mailing list
    Haskell-Cafe <at> haskell.org <mailto:Haskell-Cafe <at> haskell.org>
    http://www.haskell.org/mailman/listinfo/haskell-cafe




--
Gregory Collins <greg <at> gregorycollins.net <mailto:greg <at> gregorycollins.net>>


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe

_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe
Michael Snoyman | 9 Mar 18:05 2013

Re: File I/O benchmark help (conduit, io-streams and Handle)

Just to clarify: the problem was in fact with my code, I was not passing O_TRUNC to the open system call. Gregory's C code showed me the problem. Once I add in that option, all the different benchmarks complete in roughly the same amount of time. So given that our Haskell implementations based on Handle are just about as fast as a raw C implementation, I'd say Handle is performing very well.

Apologies if I got anyone overly concerned.


On Fri, Mar 8, 2013 at 12:36 PM, Simon Marlow <marlowsd <at> gmail.com> wrote:
1GB/s for copying a file is reasonable - it's around half the memory bandwidth, so copying the data twice would give that result (assuming no actual I/O is taking place, which is what you want because actual I/O will swamp any differences at the software level).

The Handle overhead should be negligible if you're only using hGetBufSome and hPutBuf, because those functions basically just call read() and write() when the amount of data is larger than the buffer size.

There's clearly something suspicious going on here, unfortunately I don't have time right now to investigate, but I'll keep an eye on the thread.

Cheers,
        Simon


On 08/03/13 08:36, Gregory Collins wrote:
+Simon Marlow
A couple of comments:

  * maybe we shouldn't back the file by a Handle. io-streams does this

    by default out of the box; I had a posix file interface for unix
    (guarded by CPP) for a while but decided to ditch it for simplicity.
    If your results are correct, given how slow going by Handle seems to
    be I may revisit this, I figured it would be "good enough".
  * io-streams turns Handle buffering off in withFileAsOutput. So the

    difference shouldn't be as a result of buffering. Simon: is this an
    expected result? I presume you did some Handle debugging?
  * the IO manager should not have any bearing here because file code

    doesn't actually ever use it (epoll() doesn't work for files)
  * does the difference persist when the file size gets bigger?
  * your file descriptor code doesn't handle EINTR properly, although

    you said you checked that the file copy is being done?
  * Copying a 1MB file in 1ms gives a throughput of ~1GB/s. The other

    methods have a more believable ~70MB/s throughput.

G


On Fri, Mar 8, 2013 at 7:30 AM, Michael Snoyman <michael <at> snoyman.com
<mailto:michael <at> snoyman.com>> wrote:

    Hi all,

    I'm turning to the community for some help understanding some
    benchmark results[1]. I was curious to see how the new io-streams
    would work with conduit, as it looks like a far saner low-level
    approach than Handles. In fact, the API is so simple that the entire
    wrapper is just a few lines of code[2].

    I then added in some basic file copy benchmarks, comparing
    conduit+Handle (with ResourceT or bracket), conduit+io-streams,
    straight io-streams, and lazy I/O. All approaches fell into the same
    ballpark, with conduit+bracket and conduit+io-streams taking a
    slight lead. (I haven't analyzed that enough to know if it means
    anything, however.)

    Then I decided to pull up the NoHandle code I wrote a while ago for
    conduit. This code was written initially for Windows only, to work
    around the fact that System.IO.openFile does some file locking. To
    avoid using Handles, I wrote a simple FFI wrapper exposing open,
    read, and close system calls, ported it to POSIX, and hid it behind
    a Cabal flag. Out of curiosity, I decided to expose it and include
    it in the benchmark.

    The results are extreme. I've confirmed multiple times that the copy
    algorithm is in fact copying the file, so I don't think the test
    itself is cheating somehow. But I don't know how to explain the
    massive gap. I've run this on two different systems. The results you
    see linked are from my local machine. On an EC2 instance, the gap
    was a bit smaller, but the NoHandle code was still 75% faster than
    the others.

    My initial guess is that I'm not properly tying into the IO manager,
    but I wanted to see if the community had any thoughts. The relevant
    pieces of code are [3][4][5].

    Michael

    [1] http://static.snoyman.com/streams.html
    [2]
    https://github.com/snoyberg/conduit/blob/streams/io-streams-conduit/Data/Conduit/Streams.hs
    [3]
    https://github.com/snoyberg/conduit/blob/streams/conduit/System/PosixFile.hsc
    [4]
    https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L54
    [5]
    https://github.com/snoyberg/conduit/blob/streams/conduit/Data/Conduit/Binary.hs#L167

    _______________________________________________
    Haskell-Cafe mailing list
    Haskell-Cafe <at> haskell.org <mailto:Haskell-Cafe <at> haskell.org>
    http://www.haskell.org/mailman/listinfo/haskell-cafe




--
Gregory Collins <greg <at> gregorycollins.net <mailto:greg <at> gregorycollins.net>>


_______________________________________________
Haskell-Cafe mailing list
Haskell-Cafe <at> haskell.org
http://www.haskell.org/mailman/listinfo/haskell-cafe

Gmane