[PATCH] dwarf_loader: use a better hashing function

[PATCH] dwarf_loader: use a better hashing function

[Bill Wendling]

This hashing function[1] produces better hash table bucket
distributions. The original hashing function always produced zeros in
the three least significant bits.

The new hashing funciton gives a modest performance boost.

      Original      New
       0:11.41       0:11.38
       0:11.36       0:11.34
       0:11.35       0:11.26
      -----------------------
  Avg: 0:11.373      0:11.327

for a performance improvement of 0.4%.

[1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes

Signed-off-by: Bill Wendling <morbo@google.com>
---
 hash.h | 25 ++++++++++---------------
 1 file changed, 10 insertions(+), 15 deletions(-)

diff --git a/hash.h b/hash.h
index d3aa416..ea201ab 100644
--- a/hash.h
+++ b/hash.h
@@ -33,22 +33,17 @@
 
 static inline uint64_t hash_64(const uint64_t val, const unsigned int bits)
 {
-	uint64_t hash = val;
+	uint64_t hash = val * 0x369DEA0F31A53F85UL + 0x255992D382208B61UL;
 
-	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
-	uint64_t n = hash;
-	n <<= 18;
-	hash -= n;
-	n <<= 33;
-	hash -= n;
-	n <<= 3;
-	hash += n;
-	n <<= 3;
-	hash -= n;
-	n <<= 4;
-	hash += n;
-	n <<= 2;
-	hash += n;
+	hash ^= hash >> 21;
+	hash ^= hash << 37;
+	hash ^= hash >>  4;
+
+	hash *= 0x422E19E1D95D2F0DUL;
+
+	hash ^= hash << 20;
+	hash ^= hash >> 41;
+	hash ^= hash <<  5;
 
 	/* High bits are more random, so use them. */
 	return hash >> (64 - bits);
-- 
2.30.0.478.g8a0d178c01-goog

Re: [PATCH] dwarf_loader: use a better hashing function

[Andrii Nakryiko]

On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
>
> This hashing function[1] produces better hash table bucket
> distributions. The original hashing function always produced zeros in
> the three least significant bits.
>
> The new hashing funciton gives a modest performance boost.
>
>       Original      New
>        0:11.41       0:11.38
>        0:11.36       0:11.34
>        0:11.35       0:11.26
>       -----------------------
>   Avg: 0:11.373      0:11.327
>
> for a performance improvement of 0.4%.
>
> [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes
>

Can you please also test with the one libbpf uses internally:

return (val * 11400714819323198485llu) >> (64 - bits);

?

Thanks!

> Signed-off-by: Bill Wendling <morbo@google.com>
> ---
>  hash.h | 25 ++++++++++---------------
>  1 file changed, 10 insertions(+), 15 deletions(-)
>
> diff --git a/hash.h b/hash.h
> index d3aa416..ea201ab 100644
> --- a/hash.h
> +++ b/hash.h
> @@ -33,22 +33,17 @@
>
>  static inline uint64_t hash_64(const uint64_t val, const unsigned int bits)
>  {
> -       uint64_t hash = val;
> +       uint64_t hash = val * 0x369DEA0F31A53F85UL + 0x255992D382208B61UL;
>
> -       /*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
> -       uint64_t n = hash;
> -       n <<= 18;
> -       hash -= n;
> -       n <<= 33;
> -       hash -= n;
> -       n <<= 3;
> -       hash += n;
> -       n <<= 3;
> -       hash -= n;
> -       n <<= 4;
> -       hash += n;
> -       n <<= 2;
> -       hash += n;
> +       hash ^= hash >> 21;
> +       hash ^= hash << 37;
> +       hash ^= hash >>  4;
> +
> +       hash *= 0x422E19E1D95D2F0DUL;
> +
> +       hash ^= hash << 20;
> +       hash ^= hash >> 41;
> +       hash ^= hash <<  5;
>
>         /* High bits are more random, so use them. */
>         return hash >> (64 - bits);
> --
> 2.30.0.478.g8a0d178c01-goog
>

Re: [PATCH] dwarf_loader: use a better hashing function

[Bill Wendling]

On Wed, Feb 10, 2021 at 4:00 PM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
> >
> > This hashing function[1] produces better hash table bucket
> > distributions. The original hashing function always produced zeros in
> > the three least significant bits.
> >
> > The new hashing funciton gives a modest performance boost.
> >
> >       Original      New
> >        0:11.41       0:11.38
> >        0:11.36       0:11.34
> >        0:11.35       0:11.26
> >       -----------------------
> >   Avg: 0:11.373      0:11.327
> >
> > for a performance improvement of 0.4%.
> >
> > [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes
> >
>
> Can you please also test with the one libbpf uses internally:
>
> return (val * 11400714819323198485llu) >> (64 - bits);
>
> ?
>
> Thanks!
>
It's giving me a running time of ~11.11s, which is even better. Would
you like me to submit a patch?

-bw

Re: [PATCH] dwarf_loader: use a better hashing function

[Andrii Nakryiko]

On Wed, Feb 10, 2021 at 5:24 PM Bill Wendling <morbo@google.com> wrote:
>
> On Wed, Feb 10, 2021 at 4:00 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
> >
> > On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
> > >
> > > This hashing function[1] produces better hash table bucket
> > > distributions. The original hashing function always produced zeros in
> > > the three least significant bits.
> > >
> > > The new hashing funciton gives a modest performance boost.
> > >
> > >       Original      New
> > >        0:11.41       0:11.38
> > >        0:11.36       0:11.34
> > >        0:11.35       0:11.26
> > >       -----------------------
> > >   Avg: 0:11.373      0:11.327
> > >
> > > for a performance improvement of 0.4%.
> > >
> > > [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes
> > >
> >
> > Can you please also test with the one libbpf uses internally:
> >
> > return (val * 11400714819323198485llu) >> (64 - bits);
> >
> > ?
> >
> > Thanks!
> >
> It's giving me a running time of ~11.11s, which is even better. Would
> you like me to submit a patch?

faster is better, so yeah, why not? :)

>
> -bw

Re: [PATCH] dwarf_loader: use a better hashing function

[Arnaldo Carvalho de Melo]

Em Wed, Feb 10, 2021 at 05:31:48PM -0800, Andrii Nakryiko escreveu:
> On Wed, Feb 10, 2021 at 5:24 PM Bill Wendling <morbo@google.com> wrote:
> > On Wed, Feb 10, 2021 at 4:00 PM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> > > On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
> > > > This hashing function[1] produces better hash table bucket
> > > > distributions. The original hashing function always produced zeros in
> > > > the three least significant bits.

> > > > The new hashing funciton gives a modest performance boost.

> > > >       Original      New
> > > >        0:11.41       0:11.38
> > > >        0:11.36       0:11.34
> > > >        0:11.35       0:11.26
> > > >       -----------------------
> > > >   Avg: 0:11.373      0:11.327

> > > > for a performance improvement of 0.4%.

> > > > [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes

> > > Can you please also test with the one libbpf uses internally:

> > > return (val * 11400714819323198485llu) >> (64 - bits);

> > > ?

> > It's giving me a running time of ~11.11s, which is even better. Would
> > you like me to submit a patch?

> faster is better, so yeah, why not? :)

Yeah, I agree, faster is better, please make it so :-)

- Arnaldo

Re: [PATCH] dwarf_loader: use a better hashing function

[Bill Wendling]

On Thu, Feb 11, 2021 at 5:01 AM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> Em Wed, Feb 10, 2021 at 05:31:48PM -0800, Andrii Nakryiko escreveu:
> > On Wed, Feb 10, 2021 at 5:24 PM Bill Wendling <morbo@google.com> wrote:
> > > On Wed, Feb 10, 2021 at 4:00 PM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> > > > On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
> > > > > This hashing function[1] produces better hash table bucket
> > > > > distributions. The original hashing function always produced zeros in
> > > > > the three least significant bits.
>
> > > > > The new hashing funciton gives a modest performance boost.
>
> > > > >       Original      New
> > > > >        0:11.41       0:11.38
> > > > >        0:11.36       0:11.34
> > > > >        0:11.35       0:11.26
> > > > >       -----------------------
> > > > >   Avg: 0:11.373      0:11.327
>
> > > > > for a performance improvement of 0.4%.
>
> > > > > [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes
>
> > > > Can you please also test with the one libbpf uses internally:
>
> > > > return (val * 11400714819323198485llu) >> (64 - bits);
>
> > > > ?
>
> > > It's giving me a running time of ~11.11s, which is even better. Would
> > > you like me to submit a patch?
>
> > faster is better, so yeah, why not? :)
>
> Yeah, I agree, faster is better, please make it so :-)
>
Your wish is my command! :-) Done.

-bw

Re: [PATCH] dwarf_loader: use a better hashing function

[Arnaldo Carvalho de Melo]

Em Thu, Feb 11, 2021 at 10:55:32PM -0800, Bill Wendling escreveu:
> On Thu, Feb 11, 2021 at 5:01 AM Arnaldo Carvalho de Melo
> <acme@kernel.org> wrote:
> >
> > Em Wed, Feb 10, 2021 at 05:31:48PM -0800, Andrii Nakryiko escreveu:
> > > On Wed, Feb 10, 2021 at 5:24 PM Bill Wendling <morbo@google.com> wrote:
> > > > On Wed, Feb 10, 2021 at 4:00 PM Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> > > > > On Wed, Feb 10, 2021 at 3:25 PM Bill Wendling <morbo@google.com> wrote:
> > > > > > This hashing function[1] produces better hash table bucket
> > > > > > distributions. The original hashing function always produced zeros in
> > > > > > the three least significant bits.
> >
> > > > > > The new hashing funciton gives a modest performance boost.
> >
> > > > > >       Original      New
> > > > > >        0:11.41       0:11.38
> > > > > >        0:11.36       0:11.34
> > > > > >        0:11.35       0:11.26
> > > > > >       -----------------------
> > > > > >   Avg: 0:11.373      0:11.327
> >
> > > > > > for a performance improvement of 0.4%.
> >
> > > > > > [1] From Numerical Recipes, 3rd Ed. 7.1.4 Random Hashes and Random Bytes
> >
> > > > > Can you please also test with the one libbpf uses internally:
> >
> > > > > return (val * 11400714819323198485llu) >> (64 - bits);
> >
> > > > > ?
> >
> > > > It's giving me a running time of ~11.11s, which is even better. Would
> > > > you like me to submit a patch?
> >
> > > faster is better, so yeah, why not? :)
> >
> > Yeah, I agree, faster is better, please make it so :-)
> >
> Your wish is my command! :-) Done.

Thanks, looking for the patch and applying!

No go think about something else to make it faster 8-)

- Arnaldo

[PATCH v2] dwarf_loader: use a better hashing function

[Bill Wendling]

This hashing function[1] produces better hash table bucket
distributions. The original hashing function always produced zeros in
the three least significant bits. The new hashing function gives a
modest performance boost:

  Original: 0:11.373s
  New:      0:11.110s

for a performance improvement of ~2%.

[1] From the hash function used in libbpf.

Signed-off-by: Bill Wendling <morbo@google.com>
---
 hash.h | 20 +-------------------
 1 file changed, 1 insertion(+), 19 deletions(-)

diff --git a/hash.h b/hash.h
index d3aa416..6f952c7 100644
--- a/hash.h
+++ b/hash.h
@@ -33,25 +33,7 @@
 
 static inline uint64_t hash_64(const uint64_t val, const unsigned int bits)
 {
-	uint64_t hash = val;
-
-	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
-	uint64_t n = hash;
-	n <<= 18;
-	hash -= n;
-	n <<= 33;
-	hash -= n;
-	n <<= 3;
-	hash += n;
-	n <<= 3;
-	hash -= n;
-	n <<= 4;
-	hash += n;
-	n <<= 2;
-	hash += n;
-
-	/* High bits are more random, so use them. */
-	return hash >> (64 - bits);
+	return (val * 11400714819323198485LLU) >> (64 - bits);
 }
 
 static inline uint32_t hash_32(uint32_t val, unsigned int bits)
-- 
2.30.0.478.g8a0d178c01-goog

Re: [PATCH v2] dwarf_loader: use a better hashing function

[Arnaldo Carvalho de Melo]

Em Fri, Feb 12, 2021 at 12:01:04AM -0800, Bill Wendling escreveu:
> This hashing function[1] produces better hash table bucket
> distributions. The original hashing function always produced zeros in
> the three least significant bits. The new hashing function gives a
> modest performance boost:

Some tidbits:

You forgot to CC Andrii and also to add this, which I'm doing now:

Suggested-by: Andrii Nakryiko <andrii@kernel.org>

:-)

- Arnaldo
 
>   Original: 0:11.373s
>   New:      0:11.110s
> 
> for a performance improvement of ~2%.
> 
> [1] From the hash function used in libbpf.
> 
> Signed-off-by: Bill Wendling <morbo@google.com>
> ---
>  hash.h | 20 +-------------------
>  1 file changed, 1 insertion(+), 19 deletions(-)
> 
> diff --git a/hash.h b/hash.h
> index d3aa416..6f952c7 100644
> --- a/hash.h
> +++ b/hash.h
> @@ -33,25 +33,7 @@
>  
>  static inline uint64_t hash_64(const uint64_t val, const unsigned int bits)
>  {
> -	uint64_t hash = val;
> -
> -	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
> -	uint64_t n = hash;
> -	n <<= 18;
> -	hash -= n;
> -	n <<= 33;
> -	hash -= n;
> -	n <<= 3;
> -	hash += n;
> -	n <<= 3;
> -	hash -= n;
> -	n <<= 4;
> -	hash += n;
> -	n <<= 2;
> -	hash += n;
> -
> -	/* High bits are more random, so use them. */
> -	return hash >> (64 - bits);
> +	return (val * 11400714819323198485LLU) >> (64 - bits);
>  }
>  
>  static inline uint32_t hash_32(uint32_t val, unsigned int bits)
> -- 
> 2.30.0.478.g8a0d178c01-goog
> 

-- 

- Arnaldo

Re: [PATCH v2] dwarf_loader: use a better hashing function

[Arnaldo Carvalho de Melo]

Em Fri, Feb 12, 2021 at 09:37:22AM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Fri, Feb 12, 2021 at 12:01:04AM -0800, Bill Wendling escreveu:
> > This hashing function[1] produces better hash table bucket
> > distributions. The original hashing function always produced zeros in
> > the three least significant bits. The new hashing function gives a
> > modest performance boost:
> 
> Some tidbits:
> 
> You forgot to CC Andrii and also to add this, which I'm doing now:
> 
> Suggested-by: Andrii Nakryiko <andrii@kernel.org>
> 
> :-)

See below the full cset, that will go public after some more tests here.

- Arnaldo

commit 9fecc77ed82d429fd3fe49ba275465813228e617 (HEAD -> master)
Author: Bill Wendling <morbo@google.com>
Date:   Fri Feb 12 00:01:04 2021 -0800

    dwarf_loader: Use a better hashing function, from libbpf
    
    This hashing function[1] produces better hash table bucket
    distributions. The original hashing function always produced zeros in
    the three least significant bits. The new hashing function gives a
    modest performance boost:
    
      Original: 0:11.373s
      New:      0:11.110s
    
    for a performance improvement of ~2%.
    
    [1] From the hash function used in libbpf.
    
    Committer notes:
    
    Bill found the suboptimality of the hash function being used, Andrii
    suggested using the libbpf one, which ended up being better.
    
    Signed-off-by: Bill Wendling <morbo@google.com>
    Suggested-by: Andrii Nakryiko <andrii@kernel.org>
    Cc: bpf@vger.kernel.org
    Cc: dwarves@vger.kernel.org
    Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

Re: [PATCH v2] dwarf_loader: use a better hashing function

[Bill Wendling]

On Fri, Feb 12, 2021 at 4:37 AM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> Em Fri, Feb 12, 2021 at 12:01:04AM -0800, Bill Wendling escreveu:
> > This hashing function[1] produces better hash table bucket
> > distributions. The original hashing function always produced zeros in
> > the three least significant bits. The new hashing function gives a
> > modest performance boost:
>
> Some tidbits:
>
> You forgot to CC Andrii and also to add this, which I'm doing now:
>
> Suggested-by: Andrii Nakryiko <andrii@kernel.org>
>
> :-)
>
Doh! You're right. Sorry about that. Thanks for catching it!

-bw