On important technologies: LLVM, CocoaTouch, Caching, Multi-core designs
As far as I am concerned, LLVM, CocoaTouch and memory based cache servers make up the the set of software technologies that will affect all things computing next year onwards.
LLVM is going to push code compilation and optimization to the next level. Building a new language is borderline trivial using LLVM technology. You produce the IR and the LLVM backend takes care of everything for you. I will be surprised f the 'fastest' Javascript implementation for 2009 won't be based on LLVM.
Apple's CocoaTouch is so well done, so well thought out (we are still on iPone SDK 1.0 and that speaks volumes) that it will be hard not to imagine Apple advancing and reusing the technology on, say, tablets and even making available the CocoaTouch extensions to OS X existing frameworks features set.
The ever increasing complexity of web-based services along with the rising number of users of those services and the need to sustain a user experience that depends on responding to user's actions as fast as possible, calls for the kind of tools and services that utilize intelligent RAM based caching. By caching just about everything, thanks to the ratio of reads/writes, gets/puts, resources (CPU, disks, etc) use drops by orders of magnitude while at the same time satisfying the need for a perceived fast responses to a matching requests. Developers and researchers most likely will come up with even better systems, ones that deal with cache coherency transparently, mirroring and synchronization, etc. This new realization may even render expensive, large and over-complicated systems irrelevant(e.g Oracle RDBMS). The only potential problem is the saturation of the network links, which is another class of problems researchers should look into in the near future.
On the CPUs side, everyone seems to have finally agreed that we can no longer scale vertically. We have to scale horizontally by exploiting parallelization and multi-core designs, potentially coupled with technologies such as NUMA. The Cell processor, Sun's UltraSPARC T, Intel's and AMD Multi-Core designs are build on that principles. Using all those cores('threads') efficiently, both in therms of throughput, scheduling and access to system resources is not (going to be) an easy task, but the benefits and the need to go forward justify this new approach, if not make it necessary. I personally couldn't be more excited about the possibilities availed by those architectures.
Saturday, 13 December 2008 1:33 pm
Software Rendering, Filesystems
Core i7 beats Intel IGP in DirectX 10 software rasterizer : I am very excited about the upcoming Larrabee GPU (or rather, hybrid-GPU) Intel is working on. I never found it particularly interesting to be restricted to a set of APIs for for defining and drawing scenes, as opposed to the good old days where it was all about relying on optimization techniques and clever programming to get get the most out of a pure software based rasterizer. Nowadays, at least on PCs and most game consoles, you must use either DirectX or OpenGL which provide a set of benefits (everything is taken care for you by the the GPU, the driver and the API implementation layers, etc) but also take away the fun. There is a multitude of reasons why the existing model works, but one could argue that innovation and advancement of the technology is hindered by being bound to a constrained environment and set of interfaces. I can't wait to see what near-future Carmacks, Sweeneys and Abrashes will do with the return to software rendering made possible by Larrabee and new, similar products by Nvidia and AMD.
Related references: Twilight of the GPU: an epic interview with Tim Sweeney, RAD Game Tools's advanced software rasterizer for x86, Michael Abrash, legendary x86 assembly and code optimization programmer, Software Rendering on Wikipedia, Nvidia's David Kirk on CUDA, CPUs and GPUs
Migrating to ext4 : We are looking into switching to ext4 filesystem for a few nodes on our 'testbed' environment now that ext4dev is considered stable enough to be renamed to ext4. ext3 has been sufficiently stable and performs well for our data set. Hopefully ext4 will be better in both aspects. We put XFS and ReiserFS to the test a few years ago and that didn't work out very well, though XFS, at least in paper, is impressive. Sooner or later we will need to work on our own file system, which would introduce a great number of benefits to our environment and would be fun to build.
Related references: Google File System, Lustre Filesystem
Saturday, 29 November 2008 2:03 pm
QuakeCon 07
My 'hero', John Carmack, and co. came up the kind of announcements that far outweighed the ones by Sony, Nintendo and Microsoft, combined, in the recent E3 expo, as gamesindustry.biz rightfully points out. A new game technology, a new game IP, Valve distribution deals, free, web-based, Quake III arena, xbox360 Quake, this, that and the kitchen sink. Epic Game's Unreal Engine 3 may be amazing, Crysis may look breath-taking and the recent demos of Metal Gear Solid and Killzone for the PS3 may look out of this world, but the pizza and coke powered wizard (that's JC, and, yes, its too dramatic of a metaphor ) is not to be taken lightly.
I have been scouring the Net today for videos of J.C's keynote. Alas, nowhere to be found yet. I must have watched last year's QuakeCon's keynote over 64 times. Its fundamentally cool listening to JC discussing technology, gaming and see the genius radiating out of him as he sometimes branches out to 'hard core' development issues (memory access time? procedurally generated data? .. ).
Saturday, 4 August 2007 9:30 pm
