hash_long/hash_ptr() provide really bad hashing for small hash sizes and for
cases where the lower 12 bits (Page size aligned) of the hash value are 0.
A simple modulo(prime) based hash function has way better results
across a wide range of input values. The implementation uses invers
multiplication instead of a slow division.
A futex benchmark shows better results up to a factor 10 on small hashes.
Signed-off-by: Thomas Gleixner <[email protected]>
---
include/linux/hash.h | 28 ++++++++++++++++++++++++++++
lib/Kconfig | 3 +++
lib/Makefile | 1 +
lib/hashmod.c | 44 ++++++++++++++++++++++++++++++++++++++++++++
4 files changed, 76 insertions(+)
create mode 100644 lib/hashmod.c
--- a/include/linux/hash.h
+++ b/include/linux/hash.h
@@ -83,4 +83,32 @@ static inline u32 hash32_ptr(const void
return (u32)val;
}
+struct hash_modulo {
+ unsigned int pmul;
+ unsigned int prime;
+ unsigned int mask;
+};
+
+#ifdef CONFIG_HASH_MODULO
+
+int hash_modulo_params(unsigned int hash_bits, struct hash_modulo *hm);
+
+/**
+ * hash_mod - FIXME
+ */
+static inline unsigned int hash_mod(unsigned long addr, struct hash_modulo *hm)
+{
+ u32 a, m;
+
+ if (IS_ENABLED(CONFIG_64BIT)) {
+ a = addr >> 32;
+ a ^= (unsigned int) addr;
+ } else {
+ a = addr;
+ }
+ m = ((u64)a * hm->pmul) >> 32;
+ return (a - m * hm->prime) & hm->mask;
+}
+#endif
+
#endif /* _LINUX_HASH_H */
--- a/lib/Kconfig
+++ b/lib/Kconfig
@@ -185,6 +185,9 @@ config CRC8
when they need to do cyclic redundancy check according CRC8
algorithm. Module will be called crc8.
+config HASH_MODULO
+ bool
+
config AUDIT_GENERIC
bool
depends on AUDIT && !AUDIT_ARCH
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -97,6 +97,7 @@ obj-$(CONFIG_CRC32) += crc32.o
obj-$(CONFIG_CRC7) += crc7.o
obj-$(CONFIG_LIBCRC32C) += libcrc32c.o
obj-$(CONFIG_CRC8) += crc8.o
+obj-$(CONFIG_HASH_MODULO) += hashmod.o
obj-$(CONFIG_GENERIC_ALLOCATOR) += genalloc.o
obj-$(CONFIG_842_COMPRESS) += 842/
--- /dev/null
+++ b/lib/hashmod.c
@@ -0,0 +1,44 @@
+/*
+ * Modulo based hash - Global helper functions
+ *
+ * (C) 2016 Linutronix GmbH, Thomas Gleixner
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public Licence version 2 as published by
+ * the Free Software Foundation;
+ */
+
+#include <linux/hash.h>
+#include <linux/errno,h>
+#include <linux/bug.h>
+#include <linux/kernel.h>
+
+#define hash_pmul(prime) ((unsigned int)((1ULL << 32) / prime))
+
+static const struct hash_modulo hash_modulo[] = {
+ { .prime = 3, .pmul = hash_pmul(3), .mask = 0x0003 },
+ { .prime = 7, .pmul = hash_pmul(7), .mask = 0x0007 },
+ { .prime = 13, .pmul = hash_pmul(13), .mask = 0x000f },
+ { .prime = 31, .pmul = hash_pmul(31), .mask = 0x001f },
+ { .prime = 61, .pmul = hash_pmul(61), .mask = 0x003f },
+ { .prime = 127, .pmul = hash_pmul(127), .mask = 0x007f },
+ { .prime = 251, .pmul = hash_pmul(251), .mask = 0x00ff },
+ { .prime = 509, .pmul = hash_pmul(509), .mask = 0x01ff },
+ { .prime = 1021, .pmul = hash_pmul(1021), .mask = 0x03ff },
+ { .prime = 2039, .pmul = hash_pmul(2039), .mask = 0x07ff },
+ { .prime = 4093, .pmul = hash_pmul(4093), .mask = 0x0fff },
+};
+
+/**
+ * hash_modulo_params - FIXME
+ */
+int hash_modulo_params(unsigned int hash_bits, struct hash_modulo *hm)
+{
+ hash_bits -= 2;
+
+ if (hash_bits >= ARRAY_SIZE(hash_modulo))
+ return -EINVAL;
+
+ *hm = hash_modulo[hash_bits];
+ return 0;
+}
On Thu, Apr 28, 2016 at 9:42 AM, Thomas Gleixner <[email protected]> wrote:
> hash_long/hash_ptr() provide really bad hashing for small hash sizes and for
> cases where the lower 12 bits (Page size aligned) of the hash value are 0.
Hmm.
hash_long/ptr really shouldn't care about the low bits being zero at
all, because it should mix in all the bits (using a prime multiplier
and taking the high bits).
That said, numbers rule, so clearly we need to do something. It does
strike me that we would be better off just trying to improve
hash_long().
In particular, there are people and projects that have worked on
nothing but hashing. I'm not sure we should add a new hash algorithm
even if the whole "modulo prime" sounds obviously fine in theory. For
example, your 64-bit code has obvious problems if there are common
patterns in the low and the high 32 bits.. Not a problem for something
like hash_ptr(), but it can certainly be a problem for other cases.
It would be a really good idea to have some real hard numbers on the
hashing in general, but _particularly_ so if/when we start adding new
ones. Have you tested the modulus version with SMhasher, for example?
For example, there's Thomas Wang's hash function which should cascade
all the bits.
I'd really hate to add *another* ad-hoc hash when the previous ad-hoc
hash has been shown to be bad.
Linus
On Thu, 28 Apr 2016, Linus Torvalds wrote:
>
> I'd really hate to add *another* ad-hoc hash when the previous ad-hoc
> hash has been shown to be bad.
I completely agree.
I'm not a hashing wizard and I completely failed to understand why
hash_long/ptr are so horrible for the various test cases I ran.
So my ad hoc test was to use the only hash function I truly understand. It was
state of the art in my university days :) And surprise, surprise it worked
really well.
My main focus was really to solve this futex issue which plages various people
and not to dive into hashing theory for a few weeks.
I'll try to dig up some time to analyze the hash_long failure unless someone
familiar with the problem is beating me to it.
Thanks,
tglx
On Thu, Apr 28, 2016 at 4:26 PM, Thomas Gleixner <[email protected]> wrote:
>
> I'll try to dig up some time to analyze the hash_long failure unless someone
> familiar with the problem is beating me to it.
I'm not sure you need to spend time analyzing failure: if you get bad
hashing with hash_long(), then we know that is a bad hash without
having to really try to figure out why.
It's the hashes that _look_ like they might be good hashes, but
there's not a lot of analysis behind it, that I would worry about. The
simple prime modulus _should_ be fine, but at the same time I kind of
suspect we can do better. Especially since it has two multiplications.
Looking around, there's
http://burtleburtle.net/bob/hash/integer.html
and that 32-bit "full avalanche" hash in six shifts looks like it
could be better. You wouldn't want to inline it, but the point of a
full avalanche bit mixing _should_ be that you could avoid the whole
"upper bits" part, and it should work independently of the target set
size.
So if that hash works better, it would be a pretty good replacement
option for hash_int().
There is also
https://gist.github.com/badboy/6267743
that has a 64 bit to 32 bit hash function that might be useful for
"hash_long()".
Most of the people who worry about hashes tend to have strings to
hash, not just a single word like a pointer, but there's clearly
people around who have tried to search for good hashes that really
spread out the bits.
Linus
Thomas Gleixner wrote:
> I'm not a hashing wizard and I completely failed to understand why
> hash_long/ptr are so horrible for the various test cases I ran.
It's very simple: the constants chosen are bit-sparse, *particularly*
in the least significant bits, and only 32/64 bits of the product are
kept. Using the high-word of a double-width multiply is even better,
but some machines (*cough* SPARCv9 *cough*) don't have hardware
support for that.
So what you get is:
(0x9e370001 * (x << 12)) & 0xffffffff
= (0x9e370001 * x & 0xfffff) << 12
= (0x70001 * x & 0xfffff) << 12
*Now* does it make sense?
64 bits is just as bad... 0x9e37fffffffc0001 becomes
0x7fffffffc0001, which is 2^51 - 2^18 + 1.
The challenge is the !CONFIG_ARCH_HAS_FAST_MULTIPLIER case,
when it has to be done with shifts and adds/subtracts.
Now, what's odd is that it's only relevant for 64-bit platforms, and
currently only x86 and POWER7+ have it.
SPARCv9, MIPS64, ARM64, SH64, PPC64, and IA64 all have it turned off.
Is this a bug that should be fixed?
In fact, do *any* 64-bit platforms need multiply emulation?
How many 32-bit platforms nead a multiplier that's easy for GCC to
evaluate via shifts and adds?
Generlly, by the time you've got a machine grunty enough to
need 64 bits, a multiplier is quite affordable.
Anyway, assuming there exists at least one platform that needs the
shift-and-add sequence, it's quite easy to get a higher hamming weight,
you just have to use a few more registers to save some intermediate
results.
E.g.
u64 x = val, t = val, u;
x <<= 2;
u = x += t; /* val * 5 */
x <<= 4; /* val * 80 */
x -= u; /* val * 75 = 0b1001011 */
Shall I try to come up with something?
Footnote: useful web pages on shift-and-add/subtract mutliplciation
http://www.vinc17.org/research/mulbyconst/index.en.html
http://www.spiral.net/hardware/multless.html
On Thu, Apr 28, 2016 at 7:57 PM, George Spelvin <[email protected]> wrote:
>
> How many 32-bit platforms nead a multiplier that's easy for GCC to
> evaluate via shifts and adds?
>
> Generlly, by the time you've got a machine grunty enough to
> need 64 bits, a multiplier is quite affordable.
Probably true.
That said, the whole "use a multiply to do bit shifts and adds" may be
a false economy too. It's a good trick, but it does limit the end
result in many ways: you are limited to (a) only left-shifts and (b)
only addition and subtraction.
The "only left-shifts" means that you will always be in the situation
that you'll then need to use the high bits (so you'll always need that
shift down). And being limited to just the adder tends to mean that
it's harder to get a nice spread of bits - you're basically always
going to have that same carry chain.
Having looked around at other hashes, I suspect we should look at the
ones that do five or six shifts, and a mix of add/sub and xor. And
because they shift the bits around more freely you don't have the
final shift (that ends up being dependent on the size of the target
set).
It really would be lovely to hear that we can just replace
hash_int/long() with a better hash. And I wouldn't get too hung up on
the multiplication trick. I suspect it's not worth it.
Linus
Linus wrote:
> Having looked around at other hashes, I suspect we should look at the
> ones that do five or six shifts, and a mix of add/sub and xor. And
> because they shift the bits around more freely you don't have the
> final shift (that ends up being dependent on the size of the target
> set).
I'm not sure that final shift is a problem. You need to mask the result
to the desired final size somehow, and a shift is no more cycles than
an AND.
> It really would be lovely to hear that we can just replace
> hash_int/long() with a better hash. And I wouldn't get too hung up on
> the multiplication trick. I suspect it's not worth it.
My main concern is that the scope of the search grows enormously
if we include such things. I don't want to discourage someone
from looking, but I volunteered to find a better multiplication
constant with an efficient add/subtract chain, not start a thesis
project on more general hash functions.
Two places one could look for ideas, though:
http://www.burtleburtle.net/bob/hash/integer.html
https://gist.github.com/badboy/6267743
Here's Thomas Wang's 64-bit hash, which is reputedly quite
good, in case it helps:
uint64_t hash(uint64_t key)
{
key = ~key + (key << 21); // key = (key << 21) - key - 1;
key ^= key >> 24;
key += (key << 3)) + (key << 8); // key *= 265
key ^= key >> 14;
key += (key << 2)) + (key << 4); // key *= 21
key ^= key >> 28;
key += key << 31;
return key;
}
And his slightly shorter 64-to-32-bit function:
unsigned hash(uint64_t key)
{
key = ~key + (key << 18); // key = (key << 18) - key - 1;
key ^= key >> 31;
key *= 21; // key += (key << 2)) + (key << 4);
key ^= key >> 11;
key += key << 6;
key ^= key >> 22;
return (uint32_t)key;
}
Sticking to multiplication, using the heuristics in the
current comments (prime near golden ratio = 9e3779b9 = 2654435769,)
I can come up with this for multiplying by 2654435599 = 0x9e37790f:
// -----------------------------------------------------------------------------
// This code was generated by Spiral Multiplier Block Generator, http://www.spiral.net
// Copyright (c) 2006, Carnegie Mellon University
// All rights reserved.
// The generated code is distributed under a BSD style license
// (see http://www.opensource.org/licenses/bsd-license.php)
// -----------------------------------------------
// Cost: 6 adds/subtracts 6 shifts 0 negations
// Depth: 5
// Input:
// int t0
// Outputs:
// int t1 = 2654435599 * t0
// -----------------------------------------------
t3 = shl(t0, 11); /* 2048*/
t2 = sub(t3, t0); /* 2047*/
t5 = shl(t2, 8); /* 524032*/
t4 = sub(t5, t2); /* 521985*/
t7 = shl(t0, 25); /* 33554432*/
t6 = add(t4, t7); /* 34076417*/
t9 = shl(t0, 9); /* 512*/
t8 = sub(t9, t0); /* 511*/
t11 = shl(t6, 4); /* 545222672*/
t10 = sub(t11, t6); /* 511146255*/
t12 = shl(t8, 22); /* 2143289344*/
t1 = add(t10, t12); /* 2654435599*/
Which translates into C as
uint32_t multiply(uint32_t x)
{
unsigned y = (x << 11) - x;
y -= y << 8;
y -= x << 25;
x -= x << 9;
y -= y << 4;
y -= x << 22;
return y;
}
Unfortunately, that utility bogs like hell on 64-bit constants.
On Thu, Apr 28, 2016 at 11:32 AM, Linus Torvalds
<[email protected]> wrote:
>
> hash_long/ptr really shouldn't care about the low bits being zero at
> all, because it should mix in all the bits (using a prime multiplier
> and taking the high bits).
Looking at this assertion, it doesn't actually pan out very well at all.
The 32-bit hash looks to be ok. Certainly not perfect, but not horrible.
The 64-bit hash seems to be quite horribly bad with lots of values. I
wrote a test-harness to check it out (some simple values just spread
out at a fixed stride), and the end results are *so* bad that I'm kind
of worried that I screwed up the test harness. But it gives quite
reasonable values for hash_32() and for the plain modulo case.
Now, the way my test-harness works (literally testing a lot of
equal-stride cases), the "modulo prime number" approach will
automatically look perfect. So my test-harness is pretty unfair in
that respect.
But the hash_32() function looks good when hashing into 16 bits, for
example. In that case it does a very good job of spreading things out.
When hashing into 17 bits, hash_32 still looks good, except it does
very badly for strides of 32. It starts doing worse for bigger hash
buckets and bigger strides.
But out hash_64() seems to do very badly on pretty much *any* pattern.
To the point where I started to doubt my test-program. But it really
looks like that multiplication constant is almost pessimally chosen.
For example, that _long_ range of bits set ("7fffffffc" in the middle)
is effectively just one bit set with a subtraction. And it's *right*
in that bit area that is supposed to shuffle bits 14-40 to the high
bits (which is what we actually *use*. So it effectively shuffles none
of those bits around at all, and if you have a stride of 4096, your'e
pretty much done for.
The 32-bit value is better in this regard, largely thanks to having
that low bit set. Thanks to that, the information in bits around 12-18
will stay in bits 12-18, and because we then only have 32 bits, if the
hash table is large enough, they will still be part of the bits when
we take the high bits. For the 64-bit case, bits 12-18 will never even
be relevant.
So I think that what happens here is that hash_32() is actually
somewhat reasonable. But hash_64() sucks donkey balls when there's a
lot of information in around bits 10-20 (which is exactly where a lot
of pointer bits have the *most* information thanks to alignment
issues.
Picking a new value almost at random (I say "almost", because I just
started with that 32-bit multiplicand value that mostly works and
shifted it up by 32 bits and then randomly added a few more bits to
avoid long ranges of ones and zeroes), I picked
#define GOLDEN_RATIO_PRIME_64 0x9e3700310c100d01UL
and it is *much* better in my test harness.
Of course, things like that depend on what patterns you test, But I
did have a "range of strides and hash sizes" I tried. So just for fun:
try changing GOLDEN_RATIO_PRIME_64 to that value, and see if the
absolutely _horrid_ page-aligned case goes away for you?
It really looks like those multiplication numbers were very very badly picked.
Still, that number doesn't do very well if the hash is small (say, 8
bits). But for slightly larger hash tables it seems to be doing much
better.
Linus
On Fri, Apr 29, 2016 at 9:32 PM, Linus Torvalds
<[email protected]> wrote:
wrote:
> For example, that _long_ range of bits set ("7fffffffc" in the middle)
> is effectively just one bit set with a subtraction. And it's *right*
> in that bit area that is supposed to shuffle bits 14-40 to the high bits
> (which is what we actually *use*. So it effectively shuffles none of those
> bits around at all, and if you have a stride of 4096, your'e pretty much
> done for.
Gee, I recall saying something a lot like that.
> 64 bits is just as bad... 0x9e37fffffffc0001 becomes
> 0x7fffffffc0001, which is 2^51 - 2^18 + 1.
After researching it, I think that the "high bits of a multiply" is
in fact a decent way to do such a hash. Interestingly, for a randomly
chosen odd multiplier A, the high k bits of the w-bit product A*x is a
universal hash function in the cryptographic sense. See section 2.3 of
http://arxiv.org/abs/1504.06804
One thing I note is that the advice in the comments to choose a prime
number is misquoting Knuth! Knuth says (vol. 3 section 6.4) the number
should be *relatively* prime to the word size, which for binary computers
simply means odd.
When we have a hardware multiplier, keeping the Hamming weight low is
a waste of time. When we don't, clever organization can do
better than the very naive addition/subtraction chain in the
current hash_64().
To multiply by the 32-bit constant 1640531527 = 0x61c88647 (which is
the negative of the golden ratio, so has identical distribution
properties) can be done in 6 shifts + adds, with a critical path
length of 7 operations (3 shifts + 4 adds).
#define GOLDEN_RATIO_32 0x61c88647 /* phi^2 = 1-phi */
/* Returns x * GOLDEN_RATIO_32 without a hardware multiplier */
unsigned hash_32(unsigned x)
{
unsigned y, z;
/* Path length */
y = (x << 19) + x; /* 1 shift + 1 add */
z = (x << 9) + y; /* 1 shift + 2 add */
x = (x << 23) + z; /* 1 shift + 3 add */
z = (z << 8) + y; /* 2 shift + 3 add */
x = (x << 6) - x; /* 2 shift + 4 add */
return (z << 3) + x; /* 3 shift + 4 add */
}
Finding a similarly efficient chain for the 64-bit golden ratio
0x9E3779B97F4A7C15 = 11400714819323198485
or
0x61C8864680B583EB = 7046029254386353131
is a bit of a challenge, but algorithms are known.
On Fri, Apr 29, 2016 at 2:10 PM, Linus Torvalds
<[email protected]> wrote:
> On Thu, Apr 28, 2016 at 11:32 AM, Linus Torvalds
> <[email protected]> wrote:
>>
>> hash_long/ptr really shouldn't care about the low bits being zero at
>> all, because it should mix in all the bits (using a prime multiplier
>> and taking the high bits).
>
> Looking at this assertion, it doesn't actually pan out very well at all.
>
> The 32-bit hash looks to be ok. Certainly not perfect, but not horrible.
>
> The 64-bit hash seems to be quite horribly bad with lots of values.
Ok, I have tried to research this a bit more. There's a lot of
confusion here that causes the fact that hash_64() sucks donkey balls.
The basic method for the hashing method by multiplication is fairly
sane. It's well-documented, and attributed to Knuth as the comment
above it says.
However, there's two problems in there that degrade the hash, and
particularly so for the 64-bit case.
The first is the confusion about multiplying with a prime number..
That actually makes no sense at all, and is in fact entirely wrong.
There's no reason to try to pick a prime number for the
multiplication, and I'm not seeing Knuth having ever suggested that.
Knuth's suggestion is to do the multiplication with a floating point
value A somewhere in between 0 and 1, and multiplying the integer with
it, and then just taking the fractional part and multiply it up by 'm'
and do the floor of that to get a number in the range 0..m
At no point are primes involved.
And our multiplication really does approximate that - except it's
being done in fixed-point arithmetic (so the thing you multiply with
is basically n./2**64, and the overflow is what gets rid of the
fractional part - then getting the "high bits" of the result is really
just multiplying by a power of two and taking the floor of the
result).
So the first mistake is thinking that the thing you should multiply
with should be prime. The primality matters for when you use a
division to get a modulus, which is presumably where the confusion
came from.
Now, what value 'A' you use doesn't seem to really matter much. Knuth
suggested the fractional part of the golden ratio, but I suspect that
is purely because it's an irrational number that is not near to 0 or
1. You could use anything, although since "random bit pattern" is part
of the issue, irrational numbers are a good starting point. I suspect
that with our patterns, there's actually seldom a good reason to do
lots of high-order bits, so you might as well pick something closer to
0, but whatever - it's only going to matter for the overflow part that
gets thrown away anyway.
The second mistake - and the one that actually causes the real problem
- is to believe that the "bit sparseness" is a good thing. It's not.
It's _very_ much not. If you don't mix the bits well in the
multiplication, you get exactly the problem we hit: certain bit
patternsjust will not mix up into the higher order bits.
So if you look at what the actual golden ratio representation *should* have bee:
#define GOLDEN_RATIO_32 0x9e3779b9
#define GOLDEN_RATIO_64 0x9e3779b97f4a7c16
and then compare it to the values we actually _use_ (bit-sparse primes
closeish to those values):
#define GOLDEN_RATIO_PRIME_32 0x9e370001UL
#define GOLDEN_RATIO_PRIME_64 0x9e37fffffffc0001UL
you start to see the problem. The right set of values have roughly 50%
of the bits set in a random pattern. The wrong set of values do not.
But as far as I an tell, you might as well use the fractional part of
'e' or 'pi' or just make random shit up that has a reasonable bit
distribution.
So we could use the fractional part of the golden ratio (phi):
0x9e3779b9
0x9e3779b97f4a7c16
or pi:
0x243f6a88
0x243f6a8885a308d3
or e:
0xb7e15162
0xb7e151628aed2a6b
or whatever. The constants don't have to be prime. They don't even
have to be odd, because the low bits will always be masked off anyway.
The whole "hash one integer value down to X bits" is very different in
this respect to things like string hashes, where you really do tend to
want primes (because you keep all bits).
But none of those are sparse. I think *some* amount of sparseness
might be ok if it allows people with bad CPU's to do it using
shift-and-adds, it just must not be as sparse as the current number,
the 64-bit one on particular.
There's presumably a few optimal values from a "spread bits out
evenly" standpoint, and they won't have anything to do with random
irrational constants, and will have everything to do with having nice
bitpatterns.
I'm adding Rik to the cc, because the original broken constants came
from him long long ago (they go back to 2002, originally only used for
the waitqueue hashing. Maybe he had some other input that caused him
to believe that the primeness actually mattered.
Linus
On Fri, Apr 29, 2016 at 4:31 PM, George Spelvin <[email protected]> wrote:
>
> After researching it, I think that the "high bits of a multiply" is
> in fact a decent way to do such a hash.
Our emails crossed. Yes. My test harness actually likes the
multiplication more than most of the specialized "spread out bits"
versions I've found, but only if the constants are better-chosen than
the ones we have now.
> One thing I note is that the advice in the comments to choose a prime
> number is misquoting Knuth! Knuth says (vol. 3 section 6.4) the number
> should be *relatively* prime to the word size, which for binary computers
> simply means odd.
At least for my tests, even that seems to actually be a total
non-issue. Yes, odd values *might* be better, but as mentioned in my
crossing email, it doesn't actually seem to matter for any case the
kernel cares about, since we tend to want to hash down to 10-20 bits
of data, so the least significant bit (particularly for the 64-bit
case) just doesn't matter all that much.
For the 32-bit case I suspect it's more noticeable, since we might be
using even half or more of the result.
But as mentioned: that least-order bit seems to be a *lot* less
important than the mix of the bits in the middle. Because even if your
input ends up being all zeroes in the low bits (it that worst-case
"page aligned pointers" case that Thomas had), and the hash multiplies
by an even number, you'll still end up just dropping all those zero
bits anyway.
> When we have a hardware multiplier, keeping the Hamming weight low is
> a waste of time. When we don't, clever organization can do
> better than the very naive addition/subtraction chain in the
> current hash_64().
Yeah. gcc will actually do the clever stuff for the 32-bit case, afaik.
Nothing I know of does it for the 64-bit case, which is why our
current hand-written one isn't even the smark one.
> To multiply by the 32-bit constant 1640531527 = 0x61c88647 (which is
> the negative of the golden ratio, so has identical distribution
> properties) can be done in 6 shifts + adds, with a critical path
> length of 7 operations (3 shifts + 4 adds).
So the reason we don't do this for the 32-bit case is exactly that gcc
can already do this.
If you can do the same for the 64-bit case, that might be worth it.
Linus
> At least for my tests, even that seems to actually be a total
> non-issue. Yes, odd values *might* be better, but as mentioned in my
> crossing email, it doesn't actually seem to matter for any case the
> kernel cares about, since we tend to want to hash down to 10-20 bits
> of data, so the least significant bit (particularly for the 64-bit
> case) just doesn't matter all that much.
Odd is important. If the multiplier is even, the msbit of the input
doesn't affect the hash result at all. x and (x + 0x80000000) hash to
the same value, always. That just seems like a crappy hash function.
> Yeah. gcc will actually do the clever stuff for the 32-bit case, afaik.
It's not as clever as it could be; it just does the same Booth
recoding thing, a simple series of shifts with add/subtract.
Here's the ARM code that GCC produces (9 instructions, all dependent):
mult1:
add r3, r0, r0, lsl #1
rsb r3, r0, r3, lsl #5
add r3, r3, r3, lsl #4
rsb r3, r3, r3, lsl #5
add r3, r0, r3, lsl #5
add r3, r0, r3, lsl #1
add r3, r0, r3, lsl #3
add r3, r0, r3, lsl #3
rsb r0, r0, r3, lsl #3
bx lr
versus the clever code (6 instructions, #4 and #5 could dual-issue):
mult2:
add r3, r0, r0, lsl #19
add r2, r3, r0, lsl #9
add r0, r2, r0, lsl #23
add r3, r3, r2, lsl #8
rsb r0, r0, r0, lsl #6
add r0, r0, r3, lsl #3
bx lr
On Fri, Apr 29, 2016 at 5:32 PM, George Spelvin <[email protected]> wrote:
>
> Odd is important. If the multiplier is even, the msbit of the input
> doesn't affect the hash result at all.
Fair enough. My test-set was incomplete.
>> Yeah. gcc will actually do the clever stuff for the 32-bit case, afaik.
>
> It's not as clever as it could be; it just does the same Booth
> recoding thing, a simple series of shifts with add/subtract.
Ahh. I thought gcc did the Bernstein's algorithm thing, which is
exponential in the bit size. That would have explained why it only
does it for 32-bit constants.
Not doing it for 64-bit constants makes no sense if it just uses the
trivial Booth's algorithm version.
So the odd "we don't do it for 64-bit" is apparently just an
oversight, not because gcc does something clever.
Oh well.
Linus
On Fri, 2016-04-29 at 16:51 -0700, Linus Torvalds wrote:
> There's presumably a few optimal values from a "spread bits out
> evenly" standpoint, and they won't have anything to do with random
> irrational constants, and will have everything to do with having nice
> bitpatterns.
>
> I'm adding Rik to the cc, because the original broken constants came
> from him long long ago (they go back to 2002, originally only used
> for
> the waitqueue hashing. Maybe he had some other input that caused him
> to believe that the primeness actually mattered.
I do not remember where I got that hashing algorithm and
magic constant from 14 years ago, but the changelog suggests
I got it from Chuck Lever's paper.
Chuck Lever's paper does mention that primeness "adds
certain desirable qualities", and I may have read too
much into that.
I really do not remember the "bit sparse" thing at all,
and have no idea
where that came from. Googling old email
threads about the code mostly
makes me wonder "hey, where
did that person go?"
I am all for magic numbers that work better.
--
All Rights Reversed.
> Not doing it for 64-bit constants makes no sense if it just uses the
> trivial Booth's algorithm version.
AFAICT, gcc 5 *does* optimize 64-bit multiplies by constants.
Does the belief that it doesn't date back to some really old
version?
There's still a threshold where it just punts to the multiplier.
Some examples, x86-64 (gcc 6.0.1) and aarch64 (gcc 5.3.1).
Note the difference in the multiply-by-12345 routine.
return x*9;
mul9:
leaq (%rdi,%rdi,8), %rax
ret
mul9:
add x0, x0, x0, lsl 3
ret
return x*10;
mul10:
leaq (%rdi,%rdi,4), %rax
addq %rax, %rax
ret
mul10:
lsl x1, x0, 3
add x0, x1, x0, lsl 1
ret
return x*127;
mul127:
movq %rdi, %rax
salq $7, %rax
subq %rdi, %rax
ret
mul127:
lsl x1, x0, 7
sub x0, x1, x0
ret
return x*12345;
mul12345:
imulq $12345, %rdi, %rax
ret
mul12345:
lsl x1, x0, 3
sub x1, x1, x0
lsl x1, x1, 1
sub x1, x1, x0
lsl x1, x1, 3
sub x1, x1, x0
lsl x1, x1, 3
sub x0, x1, x0
lsl x1, x0, 4
sub x0, x1, x0
ret
uint64_t y = (x << 9) - (x << 3) + x;
return x + (x << 14) - (y << 3);
mul12345_manual:
movq %rdi, %rdx
salq $14, %rax
salq $9, %rdx
addq %rdi, %rax
addq %rdi, %rdx
salq $3, %rdi
subq %rdi, %rdx
salq $3, %rdx
subq %rdx, %rax
ret
mul12345_manual:
lsl x2, x0, 9
lsl x1, x0, 14
add x2, x2, x0
add x1, x1, x0
sub x0, x2, x0, lsl 3
sub x0, x1, x0, lsl 3
ret
return x*2654435769:
mul2654435769:
movl $2654435769, %eax
imulq %rdi, %rax
ret
mul2654435769:
mov x1, 31161
movk x1, 0x9e37, lsl 16
mul x0, x0, x1
ret
The problem with variant code paths like mul12345_manual is that the
infrastructure required to determine which to use is many times larger
than the code itself. :-(
On Thu, 28 Apr 2016, Linus Torvalds wrote:
> It's the hashes that _look_ like they might be good hashes, but
> there's not a lot of analysis behind it, that I would worry about. The
> simple prime modulus _should_ be fine, but at the same time I kind of
> suspect we can do better. Especially since it has two multiplications.
>
> Looking around, there's
>
> http://burtleburtle.net/bob/hash/integer.html
>
> and that 32-bit "full avalanche" hash in six shifts looks like it
> could be better. You wouldn't want to inline it, but the point of a
> full avalanche bit mixing _should_ be that you could avoid the whole
> "upper bits" part, and it should work independently of the target set
> size.
Yes. So I tested those two:
u32 hash_64(u64 key)
{
key = ~key + (key << 18);
key ^= key >> 31;
key += (key << 2)) + (key << 4);
key ^= key >> 11;
key += key << 6;
key ^= key >> 22;
return (u32) key;
}
u32 hash_32(u32 key)
{
key = (key + 0x7ed55d16) + (key << 12);
key = (key ^ 0xc761c23c) ^ (key >> 19);
key = (key + 0x165667b1) + (key << 5);
key = (key + 0xd3a2646c) ^ (key << 9);
key = (key + 0xfd7046c5) + (key << 3);
key = (key ^ 0xb55a4f09) ^ (key >> 16);
return key;
}
They are really good and the results are similar to the simple modulo prime
hash. hash64 is slightly faster as the modulo prime as it does not have the
multiplication.
I'll send a patch to replace hash_64 and hash_32.
Text size:
x86_64 i386 arm
hash_64 88 148 128
hash_32 88 84 112
So probably slightly too large to inline.
Thanks,
tglx
On Fri, 29 Apr 2016, Linus Torvalds wrote:
> Picking a new value almost at random (I say "almost", because I just
> started with that 32-bit multiplicand value that mostly works and
> shifted it up by 32 bits and then randomly added a few more bits to
> avoid long ranges of ones and zeroes), I picked
>
> #define GOLDEN_RATIO_PRIME_64 0x9e3700310c100d01UL
>
> and it is *much* better in my test harness.
>
> Of course, things like that depend on what patterns you test, But I
> did have a "range of strides and hash sizes" I tried. So just for fun:
> try changing GOLDEN_RATIO_PRIME_64 to that value, and see if the
> absolutely _horrid_ page-aligned case goes away for you?
It solves that horrid case:
https://tglx.de/~tglx/f-ops-h64-t.png
It's faster than the shifts based version but the degradation with
hyperthreading is slightly worse.
Here for comparison the 64bit -> 32 shift version
https://tglx.de/~tglx/f-ops-wang32-t.png
FYI, that works way better than the existing shift machinery in hash_64
and the modulo prime one:
https://tglx.de/~tglx/f-ops-mod-t.png
> It really looks like those multiplication numbers were very very badly picked.
Indeed.
> Still, that number doesn't do very well if the hash is small (say, 8
> bits).
I'm still waiting for the other test to complete. Will send numbers later
today.
Thanks,
tglx
On Sat, 2016-04-30 at 15:02 +0200, Thomas Gleixner wrote:
> Yes. So I tested those two:
>
> u32 hash_64(u64 key)
> {
> key = ~key + (key << 18);
> key ^= key >> 31;
> key += (key << 2)) + (key << 4);
> key ^= key >> 11;
> key += key << 6;
> key ^= key >> 22;
> return (u32) key;
> }
>
> u32 hash_32(u32 key)
> {
> key = (key + 0x7ed55d16) + (key << 12);
> key = (key ^ 0xc761c23c) ^ (key >> 19);
> key = (key + 0x165667b1) + (key << 5);
> key = (key + 0xd3a2646c) ^ (key << 9);
> key = (key + 0xfd7046c5) + (key << 3);
> key = (key ^ 0xb55a4f09) ^ (key >> 16);
> return key;
> }
>
> They are really good and the results are similar to the simple modulo prime
> hash. hash64 is slightly faster as the modulo prime as it does not have the
> multiplication.
>
> I'll send a patch to replace hash_64 and hash_32.
>
> Text size:
> x86_64 i386 arm
> hash_64 88 148 128
> hash_32 88 84 112
>
> So probably slightly too large to inline.
I use hash_32() in net/sched/sch_fq.c, for all packets sent by Google
servers. (Note that I did _not_ use hash_ptr())
That's gazillions of packets per second, and the current multiply worked
just fine in term of hash spreading.
Are you really going to use something which looks much slower ?
u32 hash_32(u32 key)
{
key = (key + 0x7ed55d16) + (key << 12);
key = (key ^ 0xc761c23c) ^ (key >> 19);
key = (key + 0x165667b1) + (key << 5);
key = (key + 0xd3a2646c) ^ (key << 9);
key = (key + 0xfd7046c5) + (key << 3);
key = (key ^ 0xb55a4f09) ^ (key >> 16);
return key;
}
Probably having a simple multiple when ARCH_HAS_FAST_MULTIPLIER is
defined might be good enough, eventually by choosing a better
GOLDEN_RATIO_PRIME_32
On Sat, Apr 30, 2016 at 9:45 AM, Eric Dumazet <[email protected]> wrote:
>
> I use hash_32() in net/sched/sch_fq.c, for all packets sent by Google
> servers. (Note that I did _not_ use hash_ptr())
>
> That's gazillions of packets per second, and the current multiply worked
> just fine in term of hash spreading.
So hash_32() really is much better than hash_64(). I think we'll tweak
it a bit, but largely leave it alone.
The 64-bit case needs to be tweaked a _lot_.
For the 32-bit case, I like the one that George Spelvin suggested:
#define GOLDEN_RATIO_32 0x61c88647 /* phi^2 = 1-phi */
because of his slow multiplier fallback version that we could also use:
/* Returns x * GOLDEN_RATIO_32 without a hardware multiplier */
unsigned hash_32(unsigned x)
{
unsigned y, z;
/* Path length */
y = (x << 19) + x; /* 1 shift + 1 add */
z = (x << 9) + y; /* 1 shift + 2 add */
x = (x << 23) + z; /* 1 shift + 3 add */
z = (z << 8) + y; /* 2 shift + 3 add */
x = (x << 6) - x; /* 2 shift + 4 add */
return (z << 3) + x; /* 3 shift + 4 add */
}
and I don't think that we really need the several big constants with
the fancy "full cascade" function.
If you have a test-case for that sch_fq.c case, it might be a good
idea to test the above GOLDEN_RATIO_32 value, but quite frankly, I
don't see any way it would be materially different from the one we use
now. It does avoid that long series of zeroes in the low bits, but
that's actually not a huge problem for the 32-bit hash to begin with.
It's not nearly as long a series (or in the wrong bit positions) as
the 64-bit hash multiplier value had.
Also, I suspect that since you hash the kernel "struct sock" pointers,
you actually never get the kinds of really bad patterns that Thomas
had.
But maybe you use hash_32() on a pointer because you noticed that
hash_long() or hash_ptr() (which use hash_64 on 64-bit architectures,
and would have been more natural) gave worse performance?
Maybe you thought that it was the bigger multiply that caused the
performance problems? If you did performance work, I suspect it really
could have been that hash_64() did a bad job for you.
Linus
On Sat, 2016-04-30 at 10:12 -0700, Linus Torvalds wrote:
> On Sat, Apr 30, 2016 at 9:45 AM, Eric Dumazet <[email protected]> wrote:
> >
> > I use hash_32() in net/sched/sch_fq.c, for all packets sent by Google
> > servers. (Note that I did _not_ use hash_ptr())
> >
> > That's gazillions of packets per second, and the current multiply worked
> > just fine in term of hash spreading.
>
> So hash_32() really is much better than hash_64(). I think we'll tweak
> it a bit, but largely leave it alone.
>
> The 64-bit case needs to be tweaked a _lot_.
Agreed ;)
>
> For the 32-bit case, I like the one that George Spelvin suggested:
>
> #define GOLDEN_RATIO_32 0x61c88647 /* phi^2 = 1-phi */
>
> because of his slow multiplier fallback version that we could also use:
>
> /* Returns x * GOLDEN_RATIO_32 without a hardware multiplier */
> unsigned hash_32(unsigned x)
> {
> unsigned y, z;
> /* Path length */
> y = (x << 19) + x; /* 1 shift + 1 add */
> z = (x << 9) + y; /* 1 shift + 2 add */
> x = (x << 23) + z; /* 1 shift + 3 add */
> z = (z << 8) + y; /* 2 shift + 3 add */
> x = (x << 6) - x; /* 2 shift + 4 add */
> return (z << 3) + x; /* 3 shift + 4 add */
> }
>
> and I don't think that we really need the several big constants with
> the fancy "full cascade" function.
>
> If you have a test-case for that sch_fq.c case, it might be a good
> idea to test the above GOLDEN_RATIO_32 value, but quite frankly, I
> don't see any way it would be materially different from the one we use
> now. It does avoid that long series of zeroes in the low bits, but
> that's actually not a huge problem for the 32-bit hash to begin with.
> It's not nearly as long a series (or in the wrong bit positions) as
> the 64-bit hash multiplier value had.
>
> Also, I suspect that since you hash the kernel "struct sock" pointers,
> you actually never get the kinds of really bad patterns that Thomas
> had.
>
> But maybe you use hash_32() on a pointer because you noticed that
> hash_long() or hash_ptr() (which use hash_64 on 64-bit architectures,
> and would have been more natural) gave worse performance?
Not at all.
At the time I did sch_fq (for linux-3.12), hash_64() was not using a
multiply yet, but this long series of shifts/add/sub
I used hash_32() because it was simply faster on my servers.
You added this multiply in linux-3.17, and I did not noticed at that
time.
>
> Maybe you thought that it was the bigger multiply that caused the
> performance problems? If you did performance work, I suspect it really
> could have been that hash_64() did a bad job for you.
Really not ;)
I could test hash_64(), more entropy can not be bad.
Thomas Gleixner wrote:
> I'll send a patch to replace hash_64 and hash_32.
Before you do that, could we look for a way to tweak the constants
in the existing hash?
It seems the basic "take the high bits of x * K" algorithm is actually
a decent hash function if K is chosen properly, and has a significant
speed advantage on machines with half-decent multipliers.
I'm researching how to do the multiply with fewer shifts and adds on
machines that need it. (Or we could use a totally different function
in that case.)
You say that
> hash64 is slightly faster as the modulo prime as it does not have the
> multiplication.
Um... are you sure you benchmarked that right? The hash_64 code you
used (Thomas Wang's 64->32-bit hash) has a critical path consisting of 6
shifts and 7 adds. I can't believe that's faster than a single multiply.
For 1,000,000 iterations on an Ivy Bridge, the multiply is 4x
faster (5x if out of line) for me!
The constants I recommend are
#define GOLDEN_RATIO_64 0x61C8864680B583EBull
#define GOLDEN_RATIO_32 0x61C88647
rdtsc times for 1,000,000 iterations of each of the two.
(The sum of all hashes is printed to prevent dead code elimination.)
hash_64 (sum) * PHI (sum) hash_64 (sum) * PHI (sum)
17552154 a52752df 3431821 2ce5398c 17485381 a52752df 3375535 2ce5398c
17522273 a52752df 3487206 2ce5398c 17551217 a52752df 3374221 2ce5398c
17546242 a52752df 3377355 2ce5398c 17494306 a52752df 3374202 2ce5398c
17495702 a52752df 3409768 2ce5398c 17505839 a52752df 3398205 2ce5398c
17501114 a52752df 3375435 2ce5398c 17539388 a52752df 3374202 2ce5398c
And with hash_64 forced inline:
13596945 a52752df 3374931 2ce5398c 13585916 a52752df 3411107 2ce5398c
13564616 a52752df 3374928 2ce5398c 13573465 a52752df 3425160 2ce5398c
13569712 a52752df 3374915 2ce5398c 13580461 a52752df 3397773 2ce5398c
13577481 a52752df 3374912 2ce5398c 13558708 a52752df 3417456 2ce5398c
13569044 a52752df 3374909 2ce5398c 13557193 a52752df 3407912 2ce5398c
That's 3.5 cycles vs. 13.5.
(I actually have two copies of the inlined code, to show code alignment
issues.)
On a Phenom, it's worse, 4 cycles vs. 35.
35083119 a52752df 4020754 2ce5398c 35068116 a52752df 4015659 2ce5398c
35074377 a52752df 4000819 2ce5398c 35068735 a52752df 4016943 2ce5398c
35067596 a52752df 4025397 2ce5398c 35074365 a52752df 4000108 2ce5398c
35071050 a52752df 4016190 2ce5398c 35058775 a52752df 4017988 2ce5398c
35055091 a52752df 4000066 2ce5398c 35201158 a52752df 4000094 2ce5398c
My simple test code appended for anyone who cares...
#include <stdint.h>
#include <stdio.h>
/* Phi = 0x0.9E3779B97F4A7C15F... */
/* -Phi = 0x0.61C8864680B583EA1... */
#define K 0x61C8864680B583EBull
static inline uint32_t hash1(uint64_t key)
{
key = ~key + (key << 18);
key ^= key >> 31;
key += (key << 2) + (key << 4);
key ^= key >> 11;
key += key << 6;
key ^= key >> 22;
return (uint32_t)key;
}
static inline uint32_t hash2(uint64_t key)
{
return (uint32_t)(key * K >> 32);
}
static inline uint64_t rdtsc(void)
{
uint32_t lo, hi;
asm volatile("rdtsc" : "=a" (lo), "=d" (hi));
return (uint64_t)hi << 32 | lo;
}
int
main(void)
{
int i, j;
uint32_t sum, sums[20];
uint64_t start, times[20];
for (i = 0; i < 20; i += 4) {
sum = 0;
start = rdtsc();
for (j = 0; j < 1000000; j++)
sum += hash1(j+0xdeadbeef);
times[i] = rdtsc() - start;
sums[i] = sum;
sum = 0;
start = rdtsc();
for (j = 0; j < 1000000; j++)
sum += hash2(j+0xdeadbeef);
times[i+1] = rdtsc() - start;
sums[i+1] = sum;
sum = 0;
start = rdtsc();
for (j = 0; j < 1000000; j++)
sum += hash1(j+0xdeadbeef);
times[i+2] = rdtsc() - start;
sums[i+2] = sum;
sum = 0;
start = rdtsc();
for (j = 0; j < 1000000; j++)
sum += hash2(j+0xdeadbeef);
times[i+3] = rdtsc() - start;
sums[i+3] = sum;
}
for (i = 0; i < 20; i++)
printf(" %llu %08x%c",
times[i], sums[i], (~i & 3) ? ' ' : '\n');
return 0;
}
