Re: fast-import's hash table is slow

From: Jeff King <peff@peff.net>
Date: Apr 3, 2020, 12:12 UTC
Message-ID: <20200403121212.GA65799@coredump.intra.peff.net>
In-Reply-To: <38be9140-546c-e3fa-fb71-c92937094a40@web.de>

On Thu, Apr 02, 2020 at 08:40:35PM +0200, René Scharfe wrote:

Show 10 quoted lines

> > And I didn't even have to pre-size the table. This really makes me
> > wonder if there's some silly inefficiency in khash which we could be
> > addressing. Or maybe open-addressing really does lose to chaining here,
> > but I think we keep the load factor low enough that it should be a win.
> 
> Or we're just unlucky.  I tried to find the difference between khash
> with and without presizing using callgrind, but came up empty.  It did
> reveal that fast-import spends 70% of its cycles in a million memset()
> calls issued (indirectly) by git_deflate_init() which in turn is called
> by store_object() which is called from parse_and_store_blob(), though.

I think that 70% is outsized in this case because we're dumping millions of 4-byte blobs. In a real repo you'd have larger blobs, as well as actual trees and commits pulling them together.

Show 13 quoted lines

> Why is the won second when handling 1M objects not showing in its
> output?  I suspect it's because it uses its custom allocator to gather
> its data.  So I ran the test with jemalloc2 preloaded:
> 
> nr_objects   master       khash       khash+preload
> 2^20         0m5.812s     0m5.600s    0m5.604s
> 2^21         0m12.913s    0m10.884s   0m10.357s
> 2^22         0m31.257s    0m21.461s   0m21.031s
> 2^23         1m20.904s    0m40.181s   0m42.607s
> 2^24         3m59.201s    1m21.104s   1m23.814s
> 
> My measurements are noisy, but my point is simply that with a different
> allocator you'd not even have seen any slowdown when switching to khash.

Yeah, that makes sense. I still prefer the hashmap solution for its lack of pointer hackery, given that it seems to perform as well or better. I'll send a cleaned-up patch in a moment.

Show 7 quoted lines

> >  struct object_entry {
> >  	struct pack_idx_entry idx;
> > -	struct object_entry *next;
> > +	struct hashmap_entry ent;
> 
> That uses 16 bytes more memory per entry on x64 than khash would.
> That's 256MB for 2^24 objects -- not ideal, but bearable, I guess.

Isn't it 8? We're dropping the old pointer and replacing it with the "next" pointer in hashmap_entry, plus our 4-byte hash code (which likely gets padded to 8).

I think it's probably OK in practice.

Show 13 quoted lines

> > +static int object_entry_hashcmp(const void *map_data,
> > +				const struct hashmap_entry *eptr,
> > +				const struct hashmap_entry *entry_or_key,
> > +				const void *keydata)
> > +{
> > +	const struct object_id *oid = keydata;
> > +	const struct object_entry *e1, *e2;
> > +
> > +	e1 = container_of(eptr, const struct object_entry, ent);
> 
> That's nicer that the pointer alchemy in the khash conversion for sure.
> 
> But why const?  Can const change the layout of a structure?  Scary.

No, I don't think it can. I mostly copied the "const" from the other container_of() hashmap sites. I don't think it matters in practice, because we're assigning the result to a const pointer anyway. But it seems a little cleaner not to momentarily cast away the constness even inside the macro.

-Peff