Advanced Multithreading algorithm (Huff0 v0.9)

 Not being entirely satisfied by latest results from Huff0 v0.8, i've been investigating the threading performance issue mentioned in the last post. It seems that the suspect was guilty : creating threads is a really costly operation.

This was not too sensible with LZ4, thanks to its naturally large chunk size, but this is not an all-around solution. Large chunks also means granularity is impacted. For example, if a test file is divided into 3 approximately equivalent chunks, on double core systems, one core will get twice more work to do, and will become the bottleneck of the system.

In order to solve such issue, jobs must be divided into smaller tasks, so that the potentially less efficient "tail job" remain small in comparison with the rest of the job, when all cores are working full steam.

In order to work on small thread tasks, the right thing to do : keep your threads alive, and feed them regularly.

There are obviously several ways to achieve this, including "do it yourself" ones. Right now, i've been experimenting 2 well documented ones for Windows : Thread Pools, and I/O completion port.

On first look, Thread Pools seems the obvious way to go. This windows API was meant for such task : just pile up jobs, and they will get dealt with as fast as system can complete them, dispatching them over available cores.
There is just a little problem, this API is relatively recent, and unavailable to Windows XP and prior systems. Quite a deal breaker.
Or is it ? Its little cousin, QueueUserWorkItem() was created in 2000, and does exactly the same.
And its usage is damn simple. The structure is the same as CreateThread(). Just use it as a kind of "fire and forget" job queue.
It's extremely efficient. Using this API, I was able to decrease chunk size by a factor 100 with no performance degradation. Quite an achievement.

There are just 2 little problems :

1) There is no API to check when all jobs are completed. Quite a dumb (or faulty?) design. This was apparently later corrected for other languages (C#, C++, even VB), but never for C. In order to know when all segments are compressed, it is necessary to create one's own method. And things are becoming quite nasty at this stage. Since this problem is pretty frequent, many methods can be found on searching Internet, several being questionably stable hackish-styled.
The most interesting one, in my opinion, is Interlocked counter, in conjunction with event wait. Well, now the whole system is a lot more complex than it was supposed to be.

2) You get no control on system resource. No way to tell for example "use just 2 cores please". This simple API uses whatever CPU resource is available.

OK, this starts to make a lot of issues. So i investigated the second method, I/O completion port.
It sounds scary, and ... it looks scary too. On using this old API's method (available since Windows NT4!), it's obvious it was never meant to become a general-purpose job scheduler. But that's one thing it does, and very well.
Initialization is significantly more complex than Thread Pools, but once this is correctly understood, you get all the advantage without the problems. Speed is excellent, number of concurrent thread is finely tuned, and job completion check is much more natural and secure.

So here it is. This new version of Huff0 (v0.9) is using I/O completion port to divide data into much smaller chunks, resulting in more stable and faster results. A little scanning improvement was also added, for the benefit of faster compression.

Detailed performance assessment :

		Huff0 v0.8	-t2		Huff0 v0.9	-t2
	Ratio	Compress	Decoding	Ratio	Compress	Decoding
Not compressible
enwik8.7z	1.000	1.42 GB/s	1.91 GB/s	1.000	1.50 GB/s	1.94 GB/s
Hardly compressible
win98-lz4hc-lit	1.024	850 MB/s	995 MB/s	1.024	905 MB/s	995 MB/s
audio1	1.070	540 MB/s	525 MB/s	1.070	560 MB/s	530 MB/s
Distributed
enwik8-lz4hc-lit	1.290	395 MB/s	375 MB/s	1.290	400 MB/s	380 MB/s
Lightly Ordered
enwik8-lz4hc-offh	1.131	375 MB/s	350 MB/s	1.131	380 MB/s	360 MB/s
Ordered
enwik8-lz4hc-ml	2.309	375 MB/s	380 MB/s	2.309	375 MB/s	385 MB/s
Squeezed
office-lz4hc-run	3.152	330 MB/s	310 MB/s	3.152	420 MB/s	400 MB/s
enwik8-lz4hc-run	4.959	475 MB/s	430 MB/s	4.959	480 MB/s	430 MB/s
Ultra compressible
loong	278	1.49 GB/s	3.22 GB/s	278	1.98 GB/s	3.50 GB/s

The interesting bit is here : this new threading method is generic, and can be applied to any other program. So this is good news for further improvements.

Multi-threaded entropy experiment : Huff0 v0.8

 As a logical step following LZ4, i've been porting multi-threading capabilities to Huff0. For the time being, it is only available in benchmark mode. But since Huff0 is not really a compressor, rather a component for compressors, i do not plan to move it to the file interface for now.

However, there were some useful learnings during this implementation.

First, on multi-threading cost.
LZ4 has a very simple design, it was only natural to keep the large block structure and adapt it to parallel compression. With Huff0, doing the same directly results in a dramatic performance slowdown.
The reason for it : creating new threads costs some performance overhead.
In order to limit this cost, a simple solution is to limit the number of times a new thread is created; a very simple way to achieve this is to deal with larger blocks of data.
Huff0 therefore features a double-segmentation layer, with larger blocks defined at thread level, while smaller ones are dispatched for entropy adaptation purpose (i.e. compression rate is better if blocks are smaller, because the statistics are more adapted to the block).

Thinking aloud, i'm wondering if performance would be better by changing threading strategy : instead of creating a new thread for each block, what about "sleeping it", filling its tables and resume it ?
I'm not sure if it would change performance conclusion, but it's probably worth a try.
[Edit] and indeed, it is, see follow-up : http://fastcompression.blogspot.com/2011/03/advanced-multithreading-analysis-huff0.html

Second, on memory access behavior.
This part is a bit more tricky.
Making a code "multi-thread compatible" requires little more than getting rid of all form of global variables, keeping only local ones (ie dynamically allocated into the stack) or, when necessary, allocating memory in a kind of "context structure" which will be provided as part of the function parameters. This second point makes the code very much "object-oriented".
Now on the findings : i've been witnessing some performance hit on highly accessed memory slots. When, for example, a statistics or counter structure is declared, and due to its nature is updated very often, performance is better when the structure is declared as global variable, rather than local ones. (Note that declaring "static local" variables has the same effect).
I'm not sure to fully understand that one, but i've found a lead : this comment :

"Global variables are accessed via a single known address, whereas local variables are typically accessed by indexing off an address register." 

In calculating the memory address, local variables require one more fast addition operation than global ones. It sounds very little, but precise benchmarks do feel a difference.
I still have to find a way around this performance hit though, if that's possible...

You can download the new version of Huff0 here.

Reaching system speed limits

Although results from CompressionRatings benchmark were good, i was nonetheless surprised to notice that decoding speed would not improve on increasing the number of threads.

The only explanation i could come up with is that LZ4 decoder is decompressing data faster than the RAM Drive can deliver. Which is likely a correct statement, considering the in-memory benchmark results (LZ4 can decompress at an average 800MB/s per core, while RAM Drive can only deliver between 400 and 500 MB/s).

However, it would not explain why compressing with 4 threads was faster than decoding...

Here also, there is a plausible explanation : writing to RAM Drive may be slower than reading. In this case, compression has an advantage, since it writes less data.

Now, let's put that hypothesis to the test. I built a quick benchmark to measure read and write speed from a RAM Drive installed into a Windows XP box. I'm using Gavotte's Ramdisk for this test. Using another ramdisk might result in different speed, but is unlikely to dramatically change the conclusions. A different OS may also change results, but since CompressionRatings run with Windows XP, i'm mostly interested in mimicking the same conditions.

On running the benchmark, i witnessed a very stable 1190 MB/s for read operations.
On the other hand, write operations were limited at "only" 770 MB/s.
So now, that's confirmed : writing is slower than reading.

It's consistent with CompressionRatings results. Extrapolating from these figures :
if compressing at a ratio of 2:1, then the ramdrive r+w speed is limited to 525 MB/s.
On decoding the same data 1:2, the speed limit is now down to 455MB/s.
Hey, that's about the same ratio as LZ4 compression/decompression speed difference  ...