Re: LuaJIT benchmark
- From: Fredrik Widlund <fredrik.widlund@xxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Mon, 17 Mar 2014 11:30:05 +0100
>
> Sigh, the Lua program doesn't even use local variables in the
> inner loop ...
Updated the benchmark with declaring x local inside the inner loop instead
of just before, which made for some 10% improvements in small/medium sized
hashes.
Oh, and before anyone thinks this really measures hash table
> implementation quality: it's just measuring the overhead of branch
> mispredictions and cache misses for an atypical use case of hash
> tables.
Currently it's measuring different hash algorithms integer lookup
performance with a hit rate of 50%. It is a naive benchmark for now because
this is currently sufficient to be able to make a point later on.
If have some extra time I'll test some variables later on I will test
different hit rates and key types. If you have a reference to a source
where statistics around hash table use cases is investigated I'll use that
as a basis.
On Sun, Mar 16, 2014 at 11:57 PM, Mike Pall <mike-1403@xxxxxxxxxx> wrote:
> Fredrik Widlund wrote:
> > Benchmark of LuaJIT tables vs Java HashMap vs C++ unordered_map
>
> Sigh, the Lua program doesn't even use local variables in the
> inner loop ...
>
> Oh, and before anyone thinks this really measures hash table
> implementation quality: it's just measuring the overhead of branch
> mispredictions and cache misses for an atypical use case of hash
> tables.
>
> [Typical hash table accesses use a) string keys with b) a high
> temporal correlation and c) they usually have a high hit-rate or a
> high miss-rate, but it's rarely 50/50.]
>
> --Mike
>
>
Other related posts:
