5 ms·
No, it's not. A mainframe is awesome for certain classes of problems, but I don't generally think of those classes as modern. Yes, you can run ML on a mainframe
by kstrauser 7y ago
No, it's not. A mainframe is awesome for certain classes of problems, but I don't generally think of those classes as modern. Yes, you can run ML on a mainframe, but its performance isn't going to scale as broadly as throwing a thousand ephemeral virtual hosts at the problem (why virtual? because I'm not running ML 24/7 and don't want to pay for the cluster while everyone is at home asleep). Sure, maybe it can handle 12 million web requests per second - I'm skeptical about the complexity of those requests but let's roll with it - but now the computing resource is more robust than the network connections you have plugged into it. Are the pipes into your data center truly more reliable than a multi-AZ, multi-region AWS deployment?
To put it bluntly, mainframes are wonderful solutions to problems that I don't personally find interesting. (Note: personally. It's not like I don't think there's a legitimate need for the financial systems that mainframes power, it's just that I'm glad I'm not the one who has to run them.) The problems I enjoy involve horizontal scaling at levels that a mainframe just isn't going to reach, and those are the kinds of problems that most people would describe as "modern". And as I don't think mainframes are a good solution to what I consider to be modern problems, neither do I consider them to be a modern platform.
Again, not that this makes them less valuable for the problems they do solve. I don't want to run a stock exchange with its gazillion transactions (as in the database and financial kind) per second on a cluster of x86 VMs.
Edit: But you want to make them modern, make it easier to play with the damn things. I can get almost any other kind of computing tech into my house for under $1,000, but it's freaking impossible to experiment on a mainframe without selling a kidney or signing a contract in blood. If your platform's "learn more" link goes to a form read by a salesperson, you've already lost me. If I have to pull out a credit card to tinker with it, you've already lost me. Contrast with AWS, for example, which practically throws free resources at you to get you interested and invested in the platform. Where are my free IBM resources that get me a mainframe login and access to developer-useful documentation? As it stands today, if I want to learn how to operate Big Iron, I pretty much have to take a job doing it full-time. Make it easy for me to poke at one at night and on weekends if you want me to tell my boss about this cool thing you're trying to sell.
- zozbot234 7y agoExcept that mainframe-like platforms are all about vertical as opposed to horizontal scaling, particularly for transactional workloads where this kind of scaling is highly relevant. You can't horizontally scale a database (beyond trivial sharding of logically-independent partitions, perhaps with some further dependency on rarely-changing data that can thus be replicated/cached locally) without severe tradeoffs in both latency and reliability.
- kstrauser 7y agoThat's pretty much what I said. Conversely, you can't vertically scale a database (beyond throwing ever more CPUs, RAM, and RAIDed SSDs) without severe tradeoffs in electric bills and support contracts, and while still not getting the advantages of geographic redundancy. For instance, notice that none of the FAANGs are serving their websites off mainframes. Cost-per-server isn't the only reason for that. My point wasn't that mainframes aren't useful - they are! - but that they're not good solutions for a lot of the new problems people are trying to solve lately. For instance, a huge portion of fun challenges are perfectly fine with eventually consistent solutions that optimize for availability, where one single piece of reliable hardware behind a couple of redundant peering connections simply can't compete with 1,000 reasonably reliable nodes distributed globally.
- zozbot234 7y ago> For instance, notice that none of the FAANGs are serving their websites off mainframes Obviously, nobody cares if the data that makes up your FB/Twitter timeline or Google search results page is less than completely reliable. The new problems people are trying to solve with horizontal scaling and distributed computing may well be "fun" for users and developers alike, but fun is quite different from generally worthwhile.
- kstrauser 7y agoI didn't say they were unreliable. I said they were eventually consistent. ACID is a wonderful model for lots of data problems, but wholly inappropriate for lots of things where availability matters more than perfect consistency before it's reconciliation time. For instance, suppose you're reading data from a million sensors in a particle collider. You absolutely don't want to model that as a million `BEGIN; INSERT INTO readings (sensor, value) VALUES ('particle_489285', 83); COMMIT;` queries. It's probably more appropriate to shard those sensors to N servers that log the readings to local journals in realtime, then upload them to a central data warehouse later after the experiment is finished. Reliability is extremely important there, but realtime ACID-style consistency would add nothing of value. Financial transactions aren't the only "real" work out there, and there are lots of "fun" projects that are deadly serious.
- lboc 7y agoFree z/OS access for a year from IBM: https://www-01.ibm.com/events/wwe/ast/mtm/audit.nsf/enrollall https://www-01.ibm.com/events/wwe/ast/mtm/audit.nsf/enrollal... All z/OS docs are freely accessible online: https://www-01.ibm.com/servers/resourcelink/svc00100.nsf/pages/zosInternetLibrary?OpenDocument https://www-01.ibm.com/servers/resourcelink/svc00100.nsf/pag... That's just the base OS docs. Middleware and compiler docs are also free from other places at ibm.com.
- collyw 7y agoI find it interesting that the problems you describe as "modern" seem to have had the solution before the problem. I mean it's probably because we had the solution of scalable parallel computing available via the cloud that the problems started to be solved. Are the problems themselves modern?
- kstrauser 7y agoI’ve worked at a few shops doing bleeding edge stuff, so that probably colors my perceptions a lot, but you’re probably onto something. And that’s kind of the history of computing, isn’t it? Weather prediction used to look a lot like “we got a telegraph from a city 200 miles west that it was raining today, so maybe we’ll get some rain tomorrow?” to “hey, these computers are getting pretty good at math. Think we could model a cold front?” I remember playing with neural networks on my home Amiga way back when, and that was a fun toy. Now we say “here’s a petabyte of epidemic records. See anything unexpected?” Or one time someone really wished they could do math even faster than the GPU would allow, and happened to glance at their graphics card, and now we have Dogecoin. I think our industry has a great record of figuring out new ways to use things that was wholly unlike what the inventors had expected, and that’s probably what makes predicting our future directions so difficult. Like, we have ideas about problems we’d like to solve with quantum computers, but once they’re widely available I guarantee someone will realize they can also be used to make better Spotify playlists or such.
- collyw 7y agoI was kind of assuming you were some youngster until you mentioned the Amiga. Good memories there.
- kstrauser 7y agoHah, I wish! No, I'm closer to graybeard turf these days - due to the SIP, that's quite literally true. (Side note: the custom keyboard I bought last week has Amiga caps for the command keys).
- js8 7y agoI do mainframes (I manipulate SMF data and other performance data, for those in the know) and I think there are many interesting problems in the domain of business applications, actually. One is data integration. Hopefully, https://www.categoricaldata.net/ https://www.categoricaldata.net/ will make a dent into this. Another is RAD. Modern trend is to use Turing-complete languages such as Java, while in the past, a mix of application-oriented languages (COBOL on the MF) and system-oriented languages (assembler on the MF) was used. I think there is some value in having a high-level language that is more domain specific for certain applications. For example like SAS, but SAS is old. Yet I don't see anything (at least anything open source or popular enough) that would be able to replace it. SQL can do some things, but it is also getting a bit old. I will not comment much on the "modern" thing, I think the hardware is actually plenty modern (for example, IIRC z/Series processors were the first that had hardware transactional memory, and they have amazing hardware tracing capabilities), and as far as the software - I find that the older software is usually more performant, for whatever reason. There is also difference between Unix and mainframe philosophy about how to approach stuff (Unix is more permissive, mainframe is more tight), which I think has its merits as well.
- posix_me_less 7y ago> Unix is more permissive, mainframe is more tight That sounds interesting, any examples? The usual complaint about Unix is lack of fine-grain control over filesystem, but in practice this does not seem to be a problem, there are ACLs, selinux, etc. Is the situation better on mainframes?
- js8 7y agoYes, in general security is controlled through ACLs on the mainframe. Execution of system programs is also limited. But I think it goes deeper, basically lots of resources are limited or require additional security access, so if something fails, the damage it can cause (due to overuse of resources) is limited. I don't know if there is anything you can't do on Linux, except perhaps pervasive encryption, which allows to have some file operations (such as backups) running on encrypted files that the user (under which the backup or restore is running) cannot read. There are also some really old things that would be considered unsafe today, but IBM is gradually fixing and deprecating them. This is pretty much because historically, z/OS was quite customizable, and lot of user space programs (utilities and middleware, not applications) had pretty intimate hooks into the operating system. Also the auditing (not only in security) is excellent. The visibility into what the mainframe system is doing is unmatched on any other platform, I think.