Slashdot Mirror
Systems Research Is Dead?
Manoj writes "Rob Pike of Bell Labs (Yes, that [Rob Pike]) says that systems
software research is irrelevant. At a time when computing is almost the definition of innovation, research in both software and hardware at universities and much of industry is becoming insular, ossified, and irrelevant. Where is innovation? At Microsoft, mostly. Exercise: compare MS software in 1990 vs MS software today. He states that Microsoft has been working hard, and that on many (not all) dimensions, their products are superior technically. Linux is the hot new thing, but it is merely a copy of the same old stuff. And anyway, the exciting thing about Linux is the development model, and researchers contributed little to that. He states that the excitement, the art, of systems research is gone. "
Mr. Pike is just upset because his beloved innovative Plan 9 is being completely and totally eclipsed by Unix-like operating systems like Linux and FreeBSD.
He bemoans the fact that computer scientists still tend to use such crufty old tools as Unix, Emacs, and TeX, despite the fact that these tools are still very capable and that they grow more capable every day. The Novelty that Mr. Pike is so obsessed with is only useful when it constitutes an improvement over what is currently in use.
Perhaps the most absurd part of Mr. Pike's lament was his comments on Standards. He states that "With so much externally imposed structure, there is little slop left for novelty." Clearly, however this is both untrue and unfair. Mr. Pike could easily create his own little completely novel computer world. He undoubtedly has the talents. However, unless this new system allows the end user to do the things that they traditionally use their computer for what's the point? Why would I want an innovative new system that can't send email or surf the web (or run Emacs for that matter)? I wouldn't, no matter how novel it might be.
More importantly it is certainly possible to innovate within the standards that he complains about. Napster, Gnutella, and FreeNet are all innovative ways of sharing files, yet all of them rely on existing standards. Mr. Pike also totally belittles the success of scripting languages like Perl, Tcl, and Python. I certainly don't use C or C++ on my new development projects. These languages all represent innovation, but they probably aren't available for Plan 9, so they don't count.
If Plan 9 were in Linux's position you can bet that Mr. Pike would be singing a different tune, even if all Plan 9 were being used for was to run Emacs, TeX, and Netscape.
First of all, MacOS X Server has been shipping for, I believe, a year now. If you regard Jobs as being Apple, then Jobs has been shipping preemptive multitasking since before OS/2 had a GUI (which is before Windows had preemptive multitasking too), and he's been working on bringing that back to Apple since '96.
Further, Apple hasn't delivered yet because they're more ambitious. They've gone through several attempts that simply didn't live up to their standards. If you want to look at places that innovate with backwards compatibility, look at Apple -- architecture change, no problem. Complete replacement of everything from the kernel to the userland, your old programs will still work. MacOS X is really innovative; it's more modern than UNIX, more stable than Windows NT or OS/2, and easier to use than anything.
--Matthew
Get off your stupid soapbox about amiga!
;-) /usr/dict/linux.words' will do!
If anyone should Be on a Soapbox bragging about innovative OSes, it should Be someone who's talking about a high-caliBer OS. Behold the one true Multimedia Befitting OS. Benchmarking shows that this OS Belongs among the Best OSes in performance. Of course, I'm speaking on Behalf of...
darn I forgot!
Of course, the company essentially took a briBe by ditching their desktop platform. Now we have to scriBe on a emBedded version if we even want to Begin to use this OS.
(Please don't laBel me a BeOS fan, I just thought this was funny
It's amazing what: 'grep be
Asking single line questions doesn't really expand anyones understanding of any subject.
To answer your question however:
Though I am not a "purist" as I make my living as a benevolent parasite (ie tech support/analyst) for Microsoft products, I really do enjoy an have a lot of respect for Linux.
Mr. Pike has some good points, but I don't think he is completely right in every respect. First off, stable operating systems are somewhat new (linux/freebsd being among them). For all its "innovation" I'd like to see windows 2000 still run after a hard drive crash (a friend just recently told me how his hard drive crashed but bsd kept going with the processes that were still in memory... my friends, THAT is innovation).
I would also say that the idea of customizability/infite choice in OS is an innovation that even microsoft is picking up. The fact that you can tweak almost everything about an OS is relatively new (thanks to the Open Source Model).
Sure, we have a long way to go, but I think the very open nature of Open Source/Free Software encourages even more research (especially in areas of application) instead of less. You have to consider that Linux (et al) have spent thier time getting the right things done first (like stability and versatility) and now are beginning to branch out.
Yes, GNOME/KDE may be copies of prexisting things, but the wonderful thing is that they have ability to be extended in a myriad of ways by anyone (hello, Eazel). These projects are perfect examples of people accomplishing feats that were previously dismissed. Now, with the education and experience gained from overcoming the hurdles to get these things running, they can now extend those capabilities as far as thier imaginations can take them. I, personally, can't wait to see the new developments in store for KDE and GNOME. I have a feeling that they are going to prove Mr. Pike wrong.
Whenever someone claims that a field is essentially complete, they are nearly always wrong. Consider what people believed in the year 1900: Physics was considered practically a completed science -- and that was right before the explosion of modern physics.
The problem with a lot of CS programs is that they teach Computer Science. What does that mean? It means that they teach the big, complex problems of fundamental compuation and not the practical details of coding in modern languages. The Honest Truth of it is that you need both. There seems to be no middle ground right now between superficial programs that teach how to write a GUI in a certain language and leave you high and dry in the area of general knowledge of principles of programming and the theories that underly the field, and Big Science programs that work from the assumption that everyone is going to spend the rest of their life writing compilers, designing languages, and proving problems NP complete.
The program that I am in right now leans towards the latter, and there are some advantages to that. Yeah, they won't teach me Visual Basic, but frankly, there are all kinds of books out there to teach you visual basic. I learned C on my own from a book, I figure I can do it again with most other languages. True, I didn't really understand C until I started writing projects in it, but I think that is really the primary value of programs that teach language specific classes comes in. It isn't important to learn the concept of a particular language in order to know it; there are very few conceptual programming models. What makes you learn a language is using it, and playing with it. So frankly, I'd rather learnt he general principles behind various Things in the computer field (general ideas about computation and complexity, models and abstractions, network concepts etc.) and then learn what I need to know in the world on the fly than be locked into taking classes that teach me a language and not the underpinings of the field.
The problem that a lot of theory programs have, however, is that they do a poor job of making the study of theory relevant to the lives of students, and they let themselves get a little too out of date in their obsession with 'big ideas'. A widely used (and poorly written) book on Computational Theory written by a dean at my school makes it seem that the only reason to study theories of intractability is that someday, your boss might ask you to solve an NP-complete problem effeciently, and you want to be able to show that it won't work. Yeah, a lot of people I know lie awake at night worrying that their next web page design is going to be for a guy who wants a web page that solves the satisfiability problem. At the same time, my entire school has not a single class that acknowledges the existance of a graphical interface. Yes, opening and closing windows on a desktop is not a timeless model of computing, but it is the dominant paradigm for user interface at the moment, and has been for a while. There is one class on 'Human/Computer Interaction', but it deals with speach recognition and optical recognition, because those are more 'algorythmically interesting.'
Bottom line is only this: we need neither more trade schools, nor more ivory towers. Producing a lot of 'programmers' that only know how to throw up a webpage with blinking mouseovers or how to plug in values in VisBac 6.0 isn't going to advance the field very much. At the same time, where the new pressures and problems are is in the real world, the world of business and commerce. Despite the resurgance of fundamental number theory as a job skill in lite of renewed interest in cryptography, more theoreticians that don't know what the hell #include means isn't going to give much of a boost either. If anything though, I think that a lot of the imbalance right now lies in teh direction of 'Programmer in a Box/Get Rich Quick' programs that emphasize bare competancy in a single area or language as being all you need to be a succesful programmer. If only for depth of character, I think a little more is called for. In the long term, no single language or program can stay current; only the knowledge of the concepts and theories underlying the field can, but only if they seem relevant enough to keep people in the classes.
Firstly, what is a system? It is a definable entity, with a definable boundary.
Using this (loose but sufficient) definition, let us look at Linux and see how it fits in. First, of course, there's CODA. This allows you to have many computers as part of a single system, in a far more flexible way than, say, for NFS. How to get such a hybrid, networked system to work in any kind of efficient way? That's the research part.
Then, there's Beowulf. How to modify network drivers (and, if necessary, protocols) to create an efficient, high-performance distributed virtual machine? Again, sounds like research to me.
Is there more? Yes! Linux is the ideal platform to experiment with newer technologies, both where you have one physical machine and many processors, or where you have many machines networked. Because of the kernel's modular structure, you can add and remove different modules on-the-fly, testing ideas in a way that would be impossible under Windows.
Is there anything new that Linux has added, beyond the model? (Which it just borrowed from the FSF, anyway!) Yes! The aforementioned CODA is a good example. Hot-loading and hot-unloading of kernel modules is another. Linux' IPv6 stack was one of the very first for any OS, first appearing for the 2.0.20 kernel, as an extra patch.
Virtual consoles, one of the delights of many a Linux user, don't exist for DOS, Windows, or many flavours of Unix.
ReiserFS, a completely tree-based FS, is unique in it's architecture. No other FS works on that principle. Oh, did I mention it was for Linux?
I'm currently working on IGMPv3, for Linux, which will be one of the first v3 implementations for a mainstrean Unix. (I know of a few others, but not many. I doubt you'll see v3 supported in Solaris anytime in the next few weeks.)
What about other areas of Systems Research, though? Networking is dead, surely! Tell that to people working on robust Anycasting, PIMv2, queue shaping (such as CBQ, RED, ECN, RSVP, etc), any of the experimental IP protocols (IPv7 - IPv9, I think), robust MobileIP (even when ISPs are, themselves, mobile), etc.
Non-networked stuff? System Interfaces are a vital part of a system, and there's plenty of work going on there, with VR, zero-buton mice, eye-tracking, speech synthesis/recognition, OCR, etc.
Internally, there's work on automatic parallelizing of code, using Critical Path Analysis. Neural nets and genetic algorithms get a fair amount of attention, as do Artificial Life-forms. Then, there's always Quantum Computing, Optronics (such as the purely optical router), Chaos Computing (based on non-linear systems), low-temperature circuits (such as the Crusoe, which also deserves a mention as an intelligently adapting circuit, for it's ability to handle non-native instruction sets, natively), etc.
WRT kernels, we now have exokernels, microkernels, monolithic kernels and distributed kernels.
Now, you tell me that ExoPC is old-hat stuff!
Lastly, but by no means least, SETI@Home, distributed.net and Cosm are moving Systems Research forwards in a way that no other projects in the history of computing have even dreamed of. And this is supposed to be the end of the road?
(Only if "road" means a bumpy cart track, in a rickety wagon pushed by Ivory Tower professors and paid for by Big Businesses eager to NOT get results before their cash-cows were milked dry.)
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
In other words, cars are not being developed to manuever in three independant dimensions. The engine of a car (cpu of a computer) is continually being refined. The console and driving instruments (UI of an OS) are continually being refined. The fact that the car travels on land is not being altered (the fact that an OS coordinates the hardware) is not being changed.
Real innovations in computing are not coming at a systems level, but at the hardware and UI level. Case in point: POSIX.1. There have been no drastic changes in the "standard" no OS adheres to in almost 10 years. The reason? no real research into what a system must do more of to work better.
Look at databases: there are 4 critical components to a database (ACID properties). The real research in databases today are not in implementing these properties, but in OODBMS or in RDBMS multi-dimensional queries, speed optimizations, etc. No real systems work.
Most of these properties of systems can only be improved on trivially and any new innovation will be found to be damn close to the POSIX.1 spec for OSes, have ACID properties of a DB, etc.
I'll give you something to ponder: is there a better way to implement virtual memory other than spilling over into hard drive? You have to be able to index them and swap them in and out--on things called pages.
Give me a faster general purpose index (or for that matter, OS-specific purpose) than a B-tree.
Until there is a new and radical type of hardware (other than storage, input, and output) there is little need to change what and OS does or how it does it radically. Rather, as mentioned way above, the techniques used now are optimized for a particular piece of hardware and/or software.
ReiserFS may be pretty cool stuff, but it hasn't led to really new things. There is the offer that it may allow constructing data structures reasonably efficiently via "hordes of tiny files," but nobody is really using that yet, and the "research" side of that is already reasonably well-understood.
For there to be real research out of something like ReiserFS would require that people start studying different ways of constructing (say) DBMSes by using the abstractions provided by the new FS.
It isn't really systems research for someone to construct a Linux emulation system to run atop EROS; what would be innovative would be to see what kinds of cool things that may have nothing to do with UNIX as we know it can be done with it.
The problem that he doesn't comment on, which seems to be an important flip side to the notion that Microsoft is a source of innovation, is that, during the 1990s, Microsoft did an impressive job of buying up top researchers, virtually closing down major systems software research groups:
If people started doing some substantial work on exploring how to powerfully connect applications together using CORBA, that could represent some new work; unfortunately, the tools are still maturing, and the mappings to C and C++ kind of suck, at least for the purposes of generating dynamic applications.
Remember, Pike's criticisms aren't based on some vague notion that Linux is useless or bad; they are based on the notion that it's not particularly innovative, from a systems software research perspective.
If 90% of your effort represents dealing with the same old ordinary UNIX stuff, that would be largely familar to a UNIX hacker of the 1970s, then whatever you're doing can't be more than 10% innovative. Note his comment that around 90% of the effort in Plan 9, which was one of the more innovative systems of the last decade, represented efforts to honor external standards. That's a problem.
If you're not part of the solution, you're part of the precipitate.