[Haskell-cafe] Parallel weirdness

[Andrew Coppin]

Bulat Ziganshin wrote:
>
>> Weird thing #1: The first time you sort the data, it takes a few
>> seconds. The other 7 times, it takes a split second - roughly 100x 
>> faster. Wuh?
>>     
>
> this looks like disk caching effects. if data are read from disj on
> first run and from disk cache on the next runs, this only means that
> your algorithm works faster than reading its data from disk
>   

Negative. No data is ever *read* from disk, only *written* to disk. (And 
each test writes to a different file.)

The data to be sorted is generated using a trivial LCG PRNG.

> there are plenty of reasons: first, -threaded make i/o overlapped
> with calculations.

Not with -N1.

> second, parallel version may exhibit better cpu
> cache behavior - such as processing all data in cache before sending
> it back to memory
>   

Again, with -N1, it is *still* only using 1 CPU core.

>> Weird thing #4: Adding "-N2" makes *everything* slow down a few percent.
>> In particular, Task Manager shows only one CPU core in use.
>>     
>
> it's easy - your algorithm isn't really parallel.
>   

Fails to explain why the parallel version is faster than the sequential 
one (even with no parallelism), or why the sequential algorithm should 
slow down with more threads. (Surely the extra threads just sit idle?)

> there are many subtle effects making optimization much more interesting
> than using simple schemas ;)  it's why i like it so much :))
>   

Well, based on the results I've seen so far, it seems that parallelism 
is a complete waste of time because it doesn't gain you anything. And 
that doesn't make a lot of sense...