A new star has born

After the calibration was done the results of the new codec are really surprising:

The interesting part of Lizard can be seen when large compression levels for Blosc are used, specially 8 and 9. Those are mapped to compression level 41 in Lizard, which means that the LIZv1 + Huffman compression method is used. Following the documentation, this matches the compression levels of Zlib and Zstd/Brotli, and it shows.

Just for reference, here it is the performance of the LZ4 codec, from which Lizard inherits a good part of its code:

And here the performance of Zstd, which also uses Huffman coding:

So, while Lizard (or at least, the current mapping that I did for it inside Blosc) in low compression levels cannot beat the speed of LZ4 or the compression ratios of Zstd, for high compression levels it clearly beats LZ4 and Zstd speed both for compression and decompression. Most specially, it works extremely well for achieving pretty reasonable compression ratios (typically better than Zlib, albeit not as good as Zstd) at very good decompression speed and exceptional compression speed (compressing at more than the memcpy() speed, at very good ratios, oh really?).

Finally, for those wondering why I have not used the LIZv1 + Huffman compression method also for the lower compression levels in Blosc, the answer is that I obviously tried that, but for some reason, this method only performs well for large buffers, whereas for small buffers (like the ones created by low compression levels in Blosc) its performance is rather poor. I was kind of getting a similar behaviour with Zstd, where performance shines for decompressing large buffers (the difference is that Lizard can compress at tremendous speed when compared with Zstd in this scenario), so I suppose this is typical when Huffman methods are used.

Finding its place among Blosc codecs

In my previous blog, I was saying that Zstd has virtually no competitor in Write Once Read Multiple scenarios. However, I think there is still a niche for codecs that, without providing the extreme compression ratios of Zstd, they still show big enough compression muscle without loosing too much compression speed. IMO, this is a good description of how Lizard performs. However, in Blosc1 we only have slots for a couple of codecs more (but that will not be a problem for Blosc2, where much more codecs will be supported), and although I am pretty enthusiastic on adding Lizard it would be nice to gather users feedback before than that. So in case you are a Blosc user, please use the lizard branch of the C-Blosc repo (UPDATE: Lizard has been merged into the C-Blosc2 repo recently) and report back your results.

Appendix: Hardware and software used

For reference, here it is the configuration that I used for producing the plots in this blog entry.

• CPU: Intel Xeon E3-1245 v5 @ 3.50GHz (4 physical cores with hyper-threading)

• OS: Ubuntu 16.04

• Compiler: GCC 6.3.0

• C-Blosc2: 2.0.0a4.dev (2017-07-29, lizard branch)

• Lizard: 1.0.0

LZ4

First, for GCC without PGO:

Now with PGO enabled:

We can see here that, similarly to BloscLZ, although the compression speed has not improved significantly, the decompression is now reaching up to 30 GB/s, and for high compression levels, up to 20 GB/s, which is pretty good.

Zstd

First, for GCC without PGO:

Now with PGO enabled:

Wow, in this case we really can see important speedups in both compressing and decompressing. Specially interesting is the decompression case where, for the higher compression levels, Zstd can reach speeds exceeding 20 GB/s (whereas without PGO it was not able to exceed 12 GB/s) which seems a bit crazy provided the wonderful compression ratios that Zstd is able to achieve. Beyond any doubt, for Write Once Read Multiple scenarios there is no competitor for Zstd, most specially when PGO is used.

This confirms that, once again, when performance is critical for your applications, PGO should be part of your daily weaponery.

Appendix: Hardware and software used

For reference, here it is the configuration that I used for producing the plots in this blog entry.

• CPU: Intel Xeon E3-1245 v5 @ 3.50GHz (4 physical cores with hyper-threading)

• OS: Ubuntu 16.04

• Compiler: GCC 6.3.0 (using PGO)

• C-Blosc2: 2.0.0a4.dev (2017-07-11)

• LZ4 codec: 1.7.5

• Zstd codec: 1.3.0

Where we come from

Just for reference, here it is the performance of the BloscLZ codec in my server before the new tuning work:

That is the typical synthetic benchmark in Blosc, but for the plotting function in the C-Blosc2 project, the actual size of each compressed buffer is shown (and not the size of the whole dataset, as in C-Blosc1). In this case, the dataset (256 MB) is split in chunks of 4 MB, and provided that our CPU has a LLC (Last Level Cache) of 8 MB, this is sort of an optimal size for achieving maximum performance (the buffers meant for Blosc usually do not exceed 4 MB for most of its common usages).

As can be seen, performance is quite good, although compression ratios left something to be desired. Furthermore, for the maximum compression level (9), the compression ratio has a regression with respect to the previous level (8). This is not too bad, and sometimes happens in any codec, but the nice thing would be to avoid it if possible.

The new BloscLZ after fine tuning

So, after a couple of hours playing with different parameters in BloscLZ and C-Blosc2, I started to realize that the new Intel CPU performed exceedingly well when asked to compress more, to the point that high compression settings were not performing that slow in comparision with low compression ones; rather the contrary: high compression settings were operating at almost the same speed than lower ones (which was a welcome surprise indeed). Hence I tried to be set quite more aggressive parameters in BloscLZ, while trying to keep the size of internal blocks in Blosc2 below 256 KB (the typical size of L2 caches in modern CPUs). This is the result:

So the compression ratios have increased quite a bit, specially for the larger compression levels (going from less than 10x to more than 20x for this benchmark). This is courtesy of the new, more agressive compression parameters. Strikingly enough, performance has also increased in general, but specially for these large compression levels. I am not completely certain on why this is the case, but probably this new CPU architecture is much better at out-of-order execution and prefetching larger blocks of data, which benefits compressing both faster even in large buffers; similarly, I am pretty sure that improvements in compiler technology (I am using a recent GCC 6.3.0 here) is pretty important for getting faster binary code. We can also see that when using 4 threads (i.e. using all the physical cores available in our CPU at hand), BloscLZ can compress faster than a memcpy() call for most of the cases, and most specially at large compression levels, as mentioned before. Oh, and we can see that we also got rid of the regression in the compression ratio for compression level 9, which is cool.

Regarding decompression speed, we can see that the new tuning gave general speed-ups of between 10% and 20%, with no significant slowdowns in any case. All in all, quite good results indeed!

Room for more improvements? Enter PGO.

To temporary end (optimization is a never ending task) this quest for speed, I am curious about the speed that we can buy by using the PGO (Profile Guided Optimization) capability that is present in most of the modern compilers. Here I am going to use the PGO of GCC in combination with our benchmark at hand so as to provide the profile for the compiler optimizer. Here are the results when PGO is applied to the new parametrization:

So, while the speed improvement for compression is not significant (albeit a bit better), the big improvement comes in the decompression speed, where we see speeds almost reaching 50 GB/s and perhaps more interestingly, more than 35 GB/s for maximum compression level, and for first time in my life as Blosc developer, I can see the speed of decompressing with one single thread being faster than memcpy() for all the compression levels.

I wonder what the PGO technique can bring to other codecs in Blosc, but that is stuff for other blog post. At any rate, the reader is encouraged to try PGO on their own setups. I am pretty sure that she will be pleased to see nice speed improvements.

Appendix: Hardware and software used

For reference, here it is the configuration that I used for producing the plots in this blog entry.

• CPU: Intel Xeon E3-1245 v5 @ 3.50GHz (4 physical cores with hyper-threading)

• OS: Ubuntu 16.04

• Compiler: GCC 6.3.0

• C-Blosc2: 2.0.0a4.dev (2017-07-14)

• BloscLZ: 1.0.6 (2017-07-14)

A compression beast for Blosc operation

As said, Zstd is meant to achieve better compression ratios than Zlib, and this is indeed the case for many situations already. But it turns out that Zstd shines specially when faced to the kind of data that is left after the shuffle (or bitshuffle) filter passes.

As for one, here it is the typical benchmark plot for compressing with Zstd on my machine (Intel Xeon E3-1245-v5 @ 3.5GHz), side-by-side with Zlib which was the codec having the best compression ratios among all the supported inside Blosc:

As can be seen, Zstd achieves a maximum compression ratio of more than 300x for this specific dataset, which is quite a lot more than the 70x achieved by Zlib. But the coolest thing is that we are not paying a performance price for this increased compression ratio, rather the contrary, because Zstd is clearly superior (up to a 25%) in compression speed to Zlib.

But one of the most distinctive features for Blosc is its ability to decompress data very fast (sometimes faster than memcpy() as I like to remind). And look at what Zstd is able to achieve in this case:

With peak speeds larger than 10 GB/s, Zstd can decompress data more than 2x faster than Zlib peaks (~ 4 GB/s). And more importantly, when it comes to decompress data at the highest compression level, Zstd can do that about 6x faster than Zlib (~6 GB/s vs ~1 GB/s), which is a welcome surprise.

Not the fastest, but a nicely balanced one

Of course, Zstd is still far from the fastest codecs in Blosc. See for example how the internal BloscLZ codec can perform in this machine:

But nevertheless, due to its impressive balance between compression ratio and speed, Zstd is called to be one of the most attractive codecs in Blosc for the near future.

As always, all these benchmarks here were made for the specific, synthetic dataset that I am using for Blosc since the beginning (mainly for reproducibility purposes). But I am pretty sure that most of the capabilities shown here will be experienced in a large variety of datasets that Blosc is meant to tackle (in fact, it would be nice if you can share your experience by adding a comment below).

Finally, my special thanks to Yann Collet, the author of Zstd (as well as LZ4, also included in Blosc) for putting his genius at the service of the community by opening not only his code, but also his mind in his amazing series of blogs about compression: http://fastcompression.blogspot.com

Appendix: What can be expected in Blosc2

Blosc2 has support for Zstd contexts and a new way to split chunks into blocks that makes codecs go faster in general. Below you have a couple of plots on how the Blosc2/Zstd couple behaves:

As can be seen, in Blosc2 Zstd can get peaks of more than 15 GB/s, almost reaching memcpy() speed in this machine (~17 GB/s). Also, decompression speed at the highest compression ratio can scale when throwing more threads at it (a thing that Blosc1 is not able to achieve), and easily surpasses 10 GB/s. Notice that reaching such a high speed while decompressing a buffer with a really high compression ratio (~300x) is really impressing. On his part, compression speed is a bit less (25%) than in Blosc1 but still quite competitive (and on par with Zlib).

This is really exciting news to be added on top of the new planned features for Blosc2.

ARM is eating the world

ARM is an increasingly popular architecture and we can find implementation exemplars of it not only in the phones, tablets or ChromeBooks, but also acting as embedded processors, as well as in providing computing power to immensely popular Raspberry Pi's and Arduinos and even environments so apparently alien to it like High Performance Computing.

Contrarily to what has been traditional for other computer platforms, one of the most important design features for ARM is to keep energy consumption under very strict limits. Nowadays, the ARM architecture can run decently powerful CPUs where each core consumes just 600 to 750 mWatt or less.

In my opinion, it is precisely this energy efficiency what makes of ARM one of the platforms with more projection to gain ground as a general computer platform in the short future. By now, we all know that ARM allows packing more cores into a single die (e.g. your phone having more cores than your laptop, anyone?). And more cores also means more combined computing throughput (albeit a bit more difficult to program), but more importantly, more cores being able to bring data from memory at the same time. Contrarily to what one might think, having different threads transmitting data from RAM to the CPU caches provides a better utilization of memory buses, and hence, a much better global memory bandwidth. This can be seen, for example, in typical Blosc benchmarks by looking at how the bandwidth grows with the number of threads in all the dots, but specially where compression ratio equals 1 (i.e. no compression is active, so Blosc is only doing a memory copy in this case).

Blosc is getting ready for ARM

So ARM is cool indeed, but what we are doing for making it a first-class citizen? For starters, we have created a new C-Blosc2 repository that is going to act as a playground for some time and where we are going to experiment with a new range of features (those will be discussed in a later post). And this is exactly the place where we have already started implementing a NEON version of the shuffle filter.

NEON is an SIMD extension in the same spirit than SSE2 or AVX2 present in Intel/AMD offerings. NEON extension was introduced in ARMv7 architecture, and is present in most of the current high-end devices (including most of the phones and tablets floating around, including the new Raspberry Pi 2). As many of you know, leveraging SIMD in modern CPUs is key for allowing Blosc to be one of the fastest compressors around, and if we wanted to be serious about ARM, NEON support had to be here, period.

The new NEON implementation of shuffle for Blosc has been entirely made by Lucian Marc, a summer student that joined the project at the beginning of July 2015. Lucian did a terrific work on implementing the shuffle filter NEON, and during the 2-months stage he did not only that, but he also had time to do a preliminary version of the bitshuffle filter as well (not completely functional yet, but as time allows, he plans to finish that).

Some hints on the measured increase in performance

So you might be asking, how fast can perform Blosc on an ARM with NEON? Well, let's start first by showing how fast it works on a Raspberry Pi 2 (Broadcom BCM2836 ARMv7 Quad Core Processor) having NEON and running Raspbian (gcc 4.7.2). To not bore people, we are going to show just decompression speeds:

It turns out that, when using the 4 cores and low compression levels, Blosc with NEON support already shows evidence that it can equal the performance of memcpy() on ARM. This is an important fact because I did not think that ARM performance was enough to allow Blosc doing that already. I was wrong.

Okay, so Blosc using NEON can be fast, but exactly how much when compared to a shuffle implementation in pure C? Here you have the figures for the generic C shuffle:

That means that NEON can accelerate the whole decompression process between 2x and 3x, which is pretty significant, and also speaks highly about the quality of Lucian's NEON implementation.

Does that mean that we can extrapolate these figures for all ARM processors out there? Not quite. In fact, the performance of a Raspberry Pi 2 is quite mild compared with other boards. So, let's see what is the performance on a ODROID-XU3 (although it has been replaced by ODROID-XU4, the XU3 has the same processor, so we are testing a pretty powerful CPU model here). This board comes with a Samsung Exynos5422 Cortex-A15 2.0 GHz quad core and Cortex™-A7 quad core CPUs, so it is a representative of the ARM Heterogeneous Multi-Processing solution (aka big.LITTLE). Here are its figures:

So, the first thing to note is the memcpy() speed that at 1.6 GB/s, is considerably faster than the RPi2 (< 0.9 GB/s). Yeah, this is a much more capable board from a computational point of view. The second thing is that decompression speed almost doubles the memcpy() speed. Again, I was very impressed because I did not expect this range of speeds at all. ARM definitely is getting in a situation where compression can be used for an advantage, computationally speaking.

The third thing to note is a bit disappointing though: why only 3 threads appear in the plot? Well, it turns out that the benchmark suite fails miserably when using 4 threads or more. As the Raspberry setup does not suffer from this problem at all, I presume that this is more related with the board or the libraries that come with the operating system (Ubuntu 14.04). This is rather unfortunate because I was really curious to see such an ARMv7 8-core beast running at full steam using the 8 threads. At any rate, time will tell if the problem is in the board or in Blosc itself.

Just to make the benchmarks a bit more complete, let me finish this benchmark section showing the performance using the generic C code for the shuffling algorithm:

If we compare with NEON figures for the ODROID board, we can see again an increase in speed of between 2x and 4x, which is crazy amazing (sorry if I seem a bit over-enthusiastic, but again, I was not really prepared for seeing this). Again, only figures for 2 threads are in this plot because the benchmark crashes for 3 threads (this is another hint that points to the fault being outside Blosc itself and not in its NEON implementation of the shuffle filter).

At decompression speeds of 3 GB/s and ~ 2 Watt of energy consumption, the ARM platform has one of the best bandwidth/Watt ratios that you can find in the market, and this can have (and will have) profound implications on how computations will be made in the short future (as the Mont Blanc initiative is trying to demonstrate).

What to expect from ARM/Blosc in the forthcoming months

This work on supporting ARM platforms is just the beginning. As ARM processors get more spread, and most specially, faster, we will need to refine the support for ARM in Blosc.

NEON support is only a part of the game, and things like efficient handling of ARM heterogeneous architectures (big.LITTLE) or making specific tweaks for ARM cache sizes will be critical so as to make of ARM a truly first-citizen for the Blosc ecosystem.

If you have ideas on what can be improved, and most specially how, we want to learn from you :) If you want to contribute code to the project, your pull requests are very welcome too! If you like what we are doing and want to see more of this, you can also sponsor us.

How it all started

I think I began to become truly interested in compression when, back in 1992, I was installing C-News, a news server package meant to handle Usenet News articles in our university. For younger audiences, Usenet News was a very popular way to discuss about all kind of topics, but at the same time it was pretty difficult to cope with the huge amount of articles, specially because spam practices started to appear by that time. As Gene Spafford put it in 1992:

"Usenet is like a herd of performing elephants with diarrhea. Massive, difficult to redirect, awe-inspiring, entertaining, and a source of mind-boggling amounts of excrement when you least expect it."

But one thing was clear: Usenet brought massive amounts of data that had to be transmitted through the typical low-bandwidth data lines of that time: 64 Kbps shared for everyone at our university.

My mission then was to bring the Usenet News feed by making use of as low of resources as possible. Of course, one of the first things that I did was to start news transmission during the night, when everyone was warm at bed and nobody was going to complain about others stealing the precious and scarce Internet bandwidth. Another measure was to subscribe to just a selection of groups so that the transmission would end before the new day would start. And of course, I started experimenting with compression for maximizing the number of groups that we could bring to our community.

Compressing Usenet News

The most used compressor by 1992 was compress, a Unix program based on the LZW compression algorithm. But LZW had patents issues, so by that time Jean-Loup Gailly and Mark Adler started the work with gzip. At the beginning of 1993 gzip 1.0 was ready for consumption and I find it exciting not only because it was not patent-encumbered, but also because it compressed way better than the previous compress program, allowed different compression levels, and it was pretty fast too (although compress still had an advantage here, IIRC).

So I talked with the university that was providing us with the News feed and we manage to start compressing it, first with compress and then with gzip. Shortly after that, while making measurements on the new gzip improvements, I discovered that the bottleneck was in our News workstation (an HP 9000-730 with a speedy PA-7000 RISC microprocessor @ 66 MHz) being unable to decompress all the gzipped stream of subscribed news on-time. The bottleneck suddenly changed from the communication line to the CPU!

I remember spending large hours playing with different combinations of data chunk sizes and gzip compression levels, plotting the results (with the fine gnuplot) before finally coming with a combination that stroked a fair balance between available bandwidth and CPU speed, maximizing the amount of news articles hitting our university. I think this was my first realization of how compression could help bringing data faster to the system, making some processes more effective. In fact, that actually blew-up my mind and made me passionate about compression technologies for the years to come.

LZO and the explosion of compression technology

During 1996, Markus F.X.J. Oberhumer started to announce the availability of his own set of LZO compressors. These consisted in many different compressors, all of them being variations of his own compression algorithm (LZO), but tweaked to achieve either better compression ratios or compression speed. The suite was claimed to being able to achieve speeds reaching 1/3 of the memory speed of the typical Pentium-class computers available at that time. An entire set of compressors being able to approach memory speed? boy, that was a very exciting news for me.

LZO was in the back of my mind when I started my work on PyTables in August 2002 and shortly after, in May 2003, PyTables gained support for LZO. My goal was indeed to accelerate data transmission from disk to the CPU (and back), and these plots are testimonial of how beneficial LZO was for achieving that goal. Again, compression was demonstrating that it could effectively increase disk bandwidth, and not only slow internet lines.

However, although LZO was free of patent issues and fast as hell, it had a big limitation for a project like PyTables: the licensing. LZO was using the GPL license, and that prevented the inclusion of its sources in distributions without re-licensing PyTables itself as GPL, a thing that I was not willing to do (PyTables has a BSD license, as it is usual in the NumPy ecosystem). Because of that, LZO was a nice compressor to be included in GPL projects like the Linux kernel itself, but not a good fit for PyTables (although support for LZO still exists, as long as it is downloaded and installed separately).

By that time (mid 2000's) it started to appear a plethora of fast compressors with the same spirit than LZO, but with more permissive licenses (typically BSD/MIT), many of them being a nice fit for PyTables.

A new compressor for PyTables/HDF5

By 2008 it was clear that PyTables needed a compressor whose sources could be included in the PyTables tarball, so minimizing the installation requirements. For this I started considering a series of libraries and immediately devised FastLZ as a nice candidate because of its simplicity and performance. Also, FastLZ had a permissive MIT license, which was what I was looking for.

But pure FastLZ was not completely satisfactory because it was not simple enough. It had 2 compression levels that complicated the implementation quite a bit, so I decided to keep just the highest level, and then optimize certain parts of it so that speed would be acceptable. These modifications gave birth to BloscLZ, which is still being default compressor in Blosc.

But I had more ideas on what other features the new Blosc compressor should have, namely, multi-threading and an integrated shuffle filter. Multi-threading made a lot of sense by 2008 because both Intel and AMD already had a wide range of multi-core processors by then, and it was clear that the race for throwing more and more cores into systems was going to intensify. A fast compressor had to be able to use all these cores dancing around, period.

Shuffle (see slide 71 of this presentation) was the other important component of the new compressor. This algorithm relies on neighboring elements of a dataset being highly correlated to improve data compression. A shuffle filter already came as part of the HDF5 library but it was implemented in pure C, and as it had an important overhead in terms of computation, I decided to do an SIMD version using the powerful SSE2 instructions present in all Intel and AMD processors since 2003. The result is that this new shuffle implementation adds almost zero overhead compared with the compression/decompression stages.

Once all of these features were implemented, I designed a pretty comprehensive suite of tests and asked the PyTables community to help me testing the new compressor in as much systems as possible. After some iterations, we were happy when the new compressor worked flawlessly compressing and decompressing hundreds of terabytes on many different Windows and Unix boxes, both in 32-bit and 64-bit. The new beast was ready to ship.

Blosc was born

I then grabbed BloscLZ, the multi-threading support and the SSE2-powered shuffle and put it all in the same package. That also became a standalone, pure C library, with no attachments to PyTables or HDF5, so any application could make use of it. I have got the first stable version (1.0) of Blosc released by July 2010. Before this, I already introduced Blosc publicly in my EuroSciPy 2009 keynote and also made a small reference to it in an article about Starving CPUs where I stated:

"As the gap between CPU and memory speed continues to widen, I expect Blosc to improve memory-to-CPU data transmission rates over an increasing range of datasets."

And that is the thing. As CPUs are getting faster, the chances for using compression for an advantage can be applied to more and more scenarios, to the point that improving the bandwidth of main memory (RAM) is becoming possible now. And surprisingly enough, the methodology for achieving that is the same than back in the C-news ages: strike a good balance between data block sizes and compression speed, and let compression make your applications handle data faster and not only making it more compact.

When seen in perspective, it has been a long quest over the last decades. During the 90's, compression was useful to improve the bandwidth of slow internet connections. In the 2000's, it made possible accelerating disk I/O operation. In the 2010's Blosc goal is making the memory subsystem faster and whether it is able to achieve this or not will be the subject of future blogs (hint: data arrangement is critical too). But one thing is clear, achieving this (by Blosc or any other compressor out there) is just a matter of time. Such is the fate of the ever increasing gap in CPU versus memory speeds.