Re: Performance degradation when factoring out common code

older
Re: Performance degradation when...

Harendra Kumar

8 Sep 2017 8 Sep '17

8:04 a.m.

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible. -harendra On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

...

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote: the

...
second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Attachments:

attachment.html (text/html — 3.6 KB)

Show replies by date

Harendra Kumar

8 Sep 8 Sep

8:49 a.m.

New subject: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket. -harendra On 8 September 2017 at 17:34, Harendra Kumar wrote:

...

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible.

-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski
...
wrote:

...
Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Simon Peyton Jones

9:05 a.m.

New subject: Performance degradation when factoring out common code

I know that this is not an easy request, but can either of you produce a small example that demonstrates your problem? If so, please open a ticket. I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the problem unless I can reproduce it. Simon From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Harendra Kumar Sent: 08 September 2017 13:50 To: Mikolaj Konarski Cc: ghc-devs@haskell.org Subject: Re: Performance degradation when factoring out common code I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket. -harendra On 8 September 2017 at 17:34, Harendra Kumar mailto:harendra.kumar@gmail.com> wrote: Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible. -harendra On 8 September 2017 at 17:10, Mikolaj Konarski mailto:mikolaj.konarski@gmail.com> wrote: Hello, I've had a similar problem that's been fixed in 8.2.1: https://ghc.haskell.org/trac/ghc/ticket/12603 You can also use some extreme global flags, such as ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively to get most the GHC subtlety and shyness out of the way when experimenting. Good luck Mikolaj On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar mailto:harendra.kumar@gmail.com> wrote:

...

Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.orgmailto:ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmail.haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-devs&data=02%7C01%7Csimonpj%40microsoft.com%7C5ff3c69fb9d447c47b5908d4f6b832de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636404718373134824&sdata=zyHYozym6TzL61Tq5CSERjqhKlxr%2ByV0j%2FyHtxmXmVE%3D&reserved=0

Harendra Kumar

9:19 a.m.

New subject: Performance degradation when factoring out common code

I will try creating a minimal example and open a ticket for the inlining problem, the one I am sure about. -harendra On 8 September 2017 at 18:35, Simon Peyton Jones wrote:

...

*I know that this is not an easy request*, but can either of you produce a small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the problem unless I can reproduce it.

Simon

*From:* ghc-devs [mailto:ghc-devs-bounces@haskell.org] *On Behalf Of *Harendra Kumar *Sent:* 08 September 2017 13:50 *To:* Mikolaj Konarski *Cc:* ghc-devs@haskell.org *Subject:* Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible.

-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote: the

...
second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmail.haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-devs&data=02%7C01%7Csimonpj%40microsoft.com%7C5ff3c69fb9d447c47b5908d4f6b832de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636404718373134824&sdata=zyHYozym6TzL61Tq5CSERjqhKlxr%2ByV0j%2FyHtxmXmVE%3D&reserved=0

Harendra Kumar

9 Sep 9 Sep

1:05 a.m.

New subject: Performance degradation when factoring out common code

While trying to come up with a minimal example I discovered one more puzzling thing. runghc is fastest, ghc is slower, ghc with optimization is slowest. This is completely reverse of the expected order. ghc -O1 (-O2 is similar): time 15.23 ms (14.72 ms .. 15.73 ms) ghc -O0: time 3.612 ms (3.548 ms .. 3.728 ms) runghc: time 2.250 ms (2.156 ms .. 2.348 ms) I am grokking it further. Any pointers will be helpful. I understand that -O2 can sometimes be slower e.g. aggressive inlining can sometimes be counterproductive. But 4x variation is a lot and this is the case with -O1 as well which should be relatively safer than -O2 in general. Worst of all runghc is significantly faster than ghc. What's going on? -harendra On 8 September 2017 at 18:49, Harendra Kumar wrote:

...

I will try creating a minimal example and open a ticket for the inlining problem, the one I am sure about.

-harendra

On 8 September 2017 at 18:35, Simon Peyton Jones wrote:

...
*I know that this is not an easy request*, but can either of you produce a small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the problem unless I can reproduce it.

Simon

*From:* ghc-devs [mailto:ghc-devs-bounces@haskell.org] *On Behalf Of *Harendra Kumar *Sent:* 08 September 2017 13:50 *To:* Mikolaj Konarski *Cc:* ghc-devs@haskell.org *Subject:* Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible.

-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski < mikolaj.konarski@gmail.com> wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmail.haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fghc-devs&data=02%7C01%7Csimonpj%40microsoft.com%7C5ff3c69fb9d447c47b5908d4f6b832de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636404718373134824&sdata=zyHYozym6TzL61Tq5CSERjqhKlxr%2ByV0j%2FyHtxmXmVE%3D&reserved=0

Matthew Pickering

4 a.m.

New subject: Performance degradation when factoring out common code

Do you have the code? On Sat, Sep 9, 2017 at 6:05 AM, Harendra Kumar wrote:

...

While trying to come up with a minimal example I discovered one more puzzling thing. runghc is fastest, ghc is slower, ghc with optimization is slowest. This is completely reverse of the expected order.

ghc -O1 (-O2 is similar):

time 15.23 ms (14.72 ms .. 15.73 ms)

ghc -O0:

time 3.612 ms (3.548 ms .. 3.728 ms)

runghc:

time 2.250 ms (2.156 ms .. 2.348 ms)

I am grokking it further. Any pointers will be helpful. I understand that -O2 can sometimes be slower e.g. aggressive inlining can sometimes be counterproductive. But 4x variation is a lot and this is the case with -O1 as well which should be relatively safer than -O2 in general. Worst of all runghc is significantly faster than ghc. What's going on?

-harendra

On 8 September 2017 at 18:49, Harendra Kumar wrote:

...
I will try creating a minimal example and open a ticket for the inlining problem, the one I am sure about.

-harendra

On 8 September 2017 at 18:35, Simon Peyton Jones wrote:

...
I know that this is not an easy request, but can either of you produce a small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the problem unless I can reproduce it.

Simon

From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Harendra Kumar Sent: 08 September 2017 13:50 To: Mikolaj Konarski Cc: ghc-devs@haskell.org Subject: Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such problems lurking in, perhaps this is an area where some improvement looks possible.

-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Harendra Kumar

4:38 a.m.

New subject: Performance degradation when factoring out common code

The code is at: https://github.com/harendra-kumar/asyncly. The benchmark code is in "benchmark/Main.hs". The relevant function is "asyncly_basic". If you want to run it, you can use the following steps to reproduce the behavior I reported below: 1) Run "stack build" 2) Run "stack runghc benchmark/Main.hs" for runghc figures 3) Run "stack ghc benchmark/Main.hs && benchmark/Main" to compile and run normally 4) Run "stack ghc -- -O2 benchmark/Main.hs && benchmark/Main" to compile and run with -O2 flag Just look at the first benchmark (asyncly-serial), you can comment out all others if you want to. Note that the library gets compiled without any optimization flags (see the ghc options in the cabal file). So what we are seeing here is just the effect of -O2 on compiling benchmarks/Main.hs. I am also trying to isolate the problem to a minimal case. I tried removing all the INLINE pragmas in the library to make sure that I am not screwing it up by asking the compiler to inline aggressively, but that does not seem to make any difference to the situation. Let me know if you need any information from me or help in running it. There are three issues that I am trying to get answers for: 1) Why runghc is faster? It means that there is a possibility for the program to run as fast as runghc runs it. How do I get that performance or an explanation of it? 2) Why -O1/O2 degrades performance so much by 4-5x. 3) The third one is the original problem that I posted in this thread, compiler is unable to match manual inlining. It is possible that this is an issue only when -O1/O2 is used and not when -O0 is used. Thanks for the help. -harendra On 9 September 2017 at 13:30, Matthew Pickering

...

wrote:

...

Do you have the code?

...
While trying to come up with a minimal example I discovered one more puzzling thing. runghc is fastest, ghc is slower, ghc with optimization is slowest. This is completely reverse of the expected order.

ghc -O1 (-O2 is similar):

time 15.23 ms (14.72 ms .. 15.73 ms)

ghc -O0:

time 3.612 ms (3.548 ms .. 3.728 ms)

runghc:

time 2.250 ms (2.156 ms .. 2.348 ms)

I am grokking it further. Any pointers will be helpful. I understand that -O2 can sometimes be slower e.g. aggressive inlining can sometimes be counterproductive. But 4x variation is a lot and this is the case with -O1 as well which should be relatively safer than -O2 in general. Worst of all runghc is significantly faster than ghc. What's going on?

-harendra

On 8 September 2017 at 18:49, Harendra Kumar wrote:

...
I will try creating a minimal example and open a ticket for the inlining problem, the one I am sure about.

-harendra

On 8 September 2017 at 18:35, Simon Peyton Jones
...
wrote:

...
I know that this is not an easy request, but can either of you produce

a

...
small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the

...
...
...
unless I can reproduce it.

Simon

From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Harendra Kumar Sent: 08 September 2017 13:50 To: Mikolaj Konarski Cc: ghc-devs@haskell.org Subject: Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with that ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar
...
...
wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such

...
...
...
lurking in, perhaps this is an area where some improvement looks

On Sat, Sep 9, 2017 at 6:05 AM, Harendra Kumar wrote: problem problems possible.

...
...
...
-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation parameterized by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%.

inlining

...
...
does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Harendra Kumar

9:23 a.m.

New subject: Performance degradation when factoring out common code

I could pinpoint one part of the problem. Please see the ticket: https://ghc.haskell.org/trac/ghc/ticket/14208. Here is the description that I wrote in the ticket: In this particular case -O2 is 2x slower than -O0 and -O0 is 2x slower than runghc. Please see the github repo: https://github.com/harendra-kumar/ghc-perf to reproduce the issue. Readme file in the repo has instructions to reproduce. The issue seems to occur when the code is placed in a different module. When all the code is in the same module the problem does not occur. In that case -O2 is faster than -O0. However, when the code is split into two modules the performance gets inverted. Also, it does not occur always, when I tried to change the code to make it simpler for repro the problem did not occur. -harendra On 9 September 2017 at 14:08, Harendra Kumar wrote:

...

The code is at: https://github.com/harendra-kumar/asyncly. The benchmark code is in "benchmark/Main.hs". The relevant function is "asyncly_basic".

If you want to run it, you can use the following steps to reproduce the behavior I reported below:

1) Run "stack build" 2) Run "stack runghc benchmark/Main.hs" for runghc figures 3) Run "stack ghc benchmark/Main.hs && benchmark/Main" to compile and run normally 4) Run "stack ghc -- -O2 benchmark/Main.hs && benchmark/Main" to compile and run with -O2 flag

Just look at the first benchmark (asyncly-serial), you can comment out all others if you want to. Note that the library gets compiled without any optimization flags (see the ghc options in the cabal file). So what we are seeing here is just the effect of -O2 on compiling benchmarks/Main.hs.

I am also trying to isolate the problem to a minimal case. I tried removing all the INLINE pragmas in the library to make sure that I am not screwing it up by asking the compiler to inline aggressively, but that does not seem to make any difference to the situation. Let me know if you need any information from me or help in running it.

There are three issues that I am trying to get answers for:

1) Why runghc is faster? It means that there is a possibility for the program to run as fast as runghc runs it. How do I get that performance or an explanation of it?

2) Why -O1/O2 degrades performance so much by 4-5x.

3) The third one is the original problem that I posted in this thread, compiler is unable to match manual inlining. It is possible that this is an issue only when -O1/O2 is used and not when -O0 is used.

Thanks for the help.

-harendra

On 9 September 2017 at 13:30, Matthew Pickering < matthewtpickering@gmail.com> wrote:

...
Do you have the code?

...
While trying to come up with a minimal example I discovered one more puzzling thing. runghc is fastest, ghc is slower, ghc with optimization is slowest. This is completely reverse of the expected order.

ghc -O1 (-O2 is similar):

time 15.23 ms (14.72 ms .. 15.73 ms)

ghc -O0:

time 3.612 ms (3.548 ms .. 3.728 ms)

runghc:

time 2.250 ms (2.156 ms .. 2.348 ms)

I am grokking it further. Any pointers will be helpful. I understand

...
-O2 can sometimes be slower e.g. aggressive inlining can sometimes be counterproductive. But 4x variation is a lot and this is the case with -O1 as well which should be relatively safer than -O2 in general. Worst of all runghc is significantly faster than ghc. What's going on?

-harendra

On 8 September 2017 at 18:49, Harendra Kumar wrote:

...
I will try creating a minimal example and open a ticket for the

inlining

...
problem, the one I am sure about.

-harendra

On 8 September 2017 at 18:35, Simon Peyton Jones < simonpj@microsoft.com> wrote:

...
I know that this is not an easy request, but can either of you

...
...
...
small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the

...
...
...
unless I can reproduce it.

Simon

From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Harendra Kumar Sent: 08 September 2017 13:50 To: Mikolaj Konarski Cc: ghc-devs@haskell.org Subject: Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with

...
...
...
ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar < harendra.kumar@gmail.com> wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such

...
...
...
lurking in, perhaps this is an area where some improvement looks

...
...
...
-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation

On Sat, Sep 9, 2017 at 6:05 AM, Harendra Kumar wrote: that produce a problem that problems possible. parameterized

...
...
...
...
by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

Harendra Kumar

10:41 p.m.

New subject: Performance degradation when factoring out common code

Ok, I filed a ticket for the inlining issue as well - https://ghc.haskell.org/trac/ghc/ticket/14211. The reproduction test case is in the same repo on the "inlining-issue" branch, here - https://github.com/harendra-kumar/ghc-perf/tree/inlining-issue . Performance with manually inlining a function is more than 10% faster compared to factoring out code and using INLINE pragma. stack bench for compiler inlined code time 46.71 ms (45.53 ms .. 47.79 ms) stack bench --flag ghc-perf:manual for manually inlined code time 39.46 ms (38.92 ms .. 39.94 ms) -harendra On 9 September 2017 at 14:08, Harendra Kumar wrote:

...

The code is at: https://github.com/harendra-kumar/asyncly. The benchmark code is in "benchmark/Main.hs". The relevant function is "asyncly_basic".

If you want to run it, you can use the following steps to reproduce the behavior I reported below:

1) Run "stack build" 2) Run "stack runghc benchmark/Main.hs" for runghc figures 3) Run "stack ghc benchmark/Main.hs && benchmark/Main" to compile and run normally 4) Run "stack ghc -- -O2 benchmark/Main.hs && benchmark/Main" to compile and run with -O2 flag

Just look at the first benchmark (asyncly-serial), you can comment out all others if you want to. Note that the library gets compiled without any optimization flags (see the ghc options in the cabal file). So what we are seeing here is just the effect of -O2 on compiling benchmarks/Main.hs.

I am also trying to isolate the problem to a minimal case. I tried removing all the INLINE pragmas in the library to make sure that I am not screwing it up by asking the compiler to inline aggressively, but that does not seem to make any difference to the situation. Let me know if you need any information from me or help in running it.

There are three issues that I am trying to get answers for:

1) Why runghc is faster? It means that there is a possibility for the program to run as fast as runghc runs it. How do I get that performance or an explanation of it?

2) Why -O1/O2 degrades performance so much by 4-5x.

3) The third one is the original problem that I posted in this thread, compiler is unable to match manual inlining. It is possible that this is an issue only when -O1/O2 is used and not when -O0 is used.

Thanks for the help.

-harendra

On 9 September 2017 at 13:30, Matthew Pickering < matthewtpickering@gmail.com> wrote:

...
Do you have the code?

...
While trying to come up with a minimal example I discovered one more puzzling thing. runghc is fastest, ghc is slower, ghc with optimization is slowest. This is completely reverse of the expected order.

ghc -O1 (-O2 is similar):

time 15.23 ms (14.72 ms .. 15.73 ms)

ghc -O0:

time 3.612 ms (3.548 ms .. 3.728 ms)

runghc:

time 2.250 ms (2.156 ms .. 2.348 ms)

I am grokking it further. Any pointers will be helpful. I understand

...
-O2 can sometimes be slower e.g. aggressive inlining can sometimes be counterproductive. But 4x variation is a lot and this is the case with -O1 as well which should be relatively safer than -O2 in general. Worst of all runghc is significantly faster than ghc. What's going on?

-harendra

On 8 September 2017 at 18:49, Harendra Kumar wrote:

...
I will try creating a minimal example and open a ticket for the

inlining

...
problem, the one I am sure about.

-harendra

On 8 September 2017 at 18:35, Simon Peyton Jones < simonpj@microsoft.com> wrote:

...
I know that this is not an easy request, but can either of you

...
...
...
small example that demonstrates your problem? If so, please open a ticket.

I don’t like hearing about people having to use trial and error with INLINE or SPECIALISE pragmas. But I can’t even begin to solve the

...
...
...
unless I can reproduce it.

Simon

From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Harendra Kumar Sent: 08 September 2017 13:50 To: Mikolaj Konarski Cc: ghc-devs@haskell.org Subject: Re: Performance degradation when factoring out common code

I should also point out that I saw performance improvements by manually factoring out and propagating some common expressions to outer loops in performance sensitive paths. Now I have made this a habit to do this manually. Not sure if something like this has also been fixed with

...
...
...
ticket or some other ticket.

-harendra

On 8 September 2017 at 17:34, Harendra Kumar < harendra.kumar@gmail.com> wrote:

Thanks Mikolaj! I have seen some surprising behavior quite a few times recently and I was wondering whether GHC should do better. In one case I had to use SPECIALIZE very aggressively, in another version of the same code it worked well without that. I have been doing a lot of trial and error with the INLINE/NOINLINE pragmas to figure out what the right combination is. Sometimes it just feels like black magic, because I cannot find a rationale to explain the behavior. I am not sure if there are any more such

...
...
...
lurking in, perhaps this is an area where some improvement looks

...
...
...
-harendra

On 8 September 2017 at 17:10, Mikolaj Konarski wrote:

Hello,

I've had a similar problem that's been fixed in 8.2.1:

https://ghc.haskell.org/trac/ghc/ticket/12603

You can also use some extreme global flags, such as

ghc-options: -fexpose-all-unfoldings -fspecialise-aggressively

to get most the GHC subtlety and shyness out of the way when experimenting.

Good luck Mikolaj

On Fri, Sep 8, 2017 at 11:21 AM, Harendra Kumar wrote:

...
Hi,

I have this code snippet for the bind implementation of a Monad:

AsyncT m >>= f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a <> (r >>= f) in m Nothing stp yield

I want to have multiple versions of this implementation

On Sat, Sep 9, 2017 at 6:05 AM, Harendra Kumar wrote: that produce a problem that problems possible. parameterized

...
...
...
...
by a function, like this:

bindWith k (AsyncT m) f = AsyncT $ \_ stp yld -> let run x = (runAsyncT x) Nothing stp yld yield a _ Nothing = run $ f a yield a _ (Just r) = run $ f a `k` (bindWith k r f) in m Nothing stp yield

And then the bind function becomes:

(>>=) = bindWith (<>)

But this leads to a performance degradation of more than 10%. inlining does not help, I tried INLINE pragma as well as the "inline" GHC builtin. I thought this should be a more or less straightforward replacement making the second version equivalent to the first one. But apparently there is something going on here that makes it perform worse.

I did not look at the core, stg or asm yet. Hoping someone can quickly comment on it. Any ideas why is it so? Can this be worked around somehow?

Thanks, Harendra

...
_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs

2811

Age (days ago)

2813

Last active (days ago)

List overview

Download

8 comments

3 participants

participants (3)

Harendra Kumar
Matthew Pickering
Simon Peyton Jones

Re: Performance degradation when factoring out common code

tags

participants (3)