Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Simple, correct, fast: in that order

drewdevault.com

Post reply on HN
Print view View original

311–320 of 349 posts

Re: Simple, correct, fast: in that order

#311

Post by Chyzwar »

Simple is the property of microscale. You can have a class, function, module that is simple but the whole system can be incredibly complex. You can sacrifice simplicity in many places as long as things are loosely coupled.

Correctness can be only achieved if the programmer has a good understanding of requirements and posses necessary discipline to write tests. I cannot image system that maintains correctness without tests.

Performance usually is something that only good architecture can bring. I disagree with the article. If you focus on simple too much you will miss important requirements and you will make architecture choices that negatively impact performance. More annoying it the author is trying to create another silver bullet approach.

Simple/Correct/Fast - it depends on the problem you try to solve.

Re: Simple, correct, fast: in that order

#313

Post by aidenn0 »

zzzcpan wrote: post #173
aidenn0 wrote: post #168

Earlier quoted context omitted.

> Correctness is the destination of any piece of software "Good enough" is the destination of any piece of software. Sometimes that means correct, but more often it means "oh yeah, sometimes it starts acting funny, just restart it when that happens"

It never means "correct". Not to mention that 100% correctness is even impossible.

They said "destination" being correct, with me interpreting "destination" in the sense of "goal." My point was that some software has the goal of being 100% correct, but most software does not.

Re: Simple, correct, fast: in that order

#314

Post by Const-me »

a1369209993 wrote: post #265
whack wrote: post #145

Earlier quoted context omitted.

And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you insist on only hiring chauffeurs who drive at 100mph, you can hardly complain when they get into a few accidents.

> And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you can find a CPU that has the same number of non-cache[1] transistors as a Intel/AMD chip, but spends them on a larger number of simple (and preferably independent/non-hyperthreaded) cores, rather than squandering them on speculative execution and ten thousand obscure model specific registers, I wo…

Intel makes them and you can buy them today, with up to 72 Atom CPU cores, e.g. (1) https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processo...

Very niche products.

For massively parallel number crunching, GPUs are much better in both performance/watt and performance/dollar. That Xeon Phi 7290 delivers up to 3.45TFlops, costs $3200, and consumes 245W. Compare with GeForce 1080Ti 10.6 TFlops, $700, same 250W.

For general purpose software they don’t work particularly well either. Most IO interfaces is serial, SATA, PCI-X, they have very few wires going to CPU. If you’re IO bound and you don’t have enough single-thread performance you’ll struggle to saturate the bandwidth, doable but very hard.

Also for general-purpose software latency matters. Namely, input to screen latency for desktops and mobiles, or request to response latency for servers. Get Windows or Android tablet with Intel Atom Z8300 (available for $80-100), and see how it performs, it has 4 very similar cores (minus AVX-512), and frequencies are very similar, too.

Re: Simple, correct, fast: in that order

#315

Post by taneq »

kungtotte wrote: post #272
AstralStorm wrote: post #242

Earlier quoted context omitted.

4) After thinking how to get beat bang for buck when optimizing. Profilers don't always pinpoint the culprit.

I have a hard time seeing how profilers don't at least point you in the right direction. Or are we using different definitions of the word? They tell you how much time is spent in any given section of your program. If 80% of the time is spent in 20% of the code, it's usually a safe bet that's where you should start looking!

Generally I agree, but after you've picked off a few low hanging fruit, you'll end up with the profiler pointing at the function which does most of your grunt work, at which point you'd expect that to take up most of the time.

Once you've got this pretty optimized and it's still taking up the lion's share of your execution time, you have to look elsewhere (probably changing your overall approach or applying some higher level optimisation) to improve things further.

Re: Simple, correct, fast: in that order

#316

Post by Veedrac »

In the boardroom, three of Ford's top brass are in the midst of heated debate...

"As you see, our main priorities, in order of importance, are simplicity of design, reliable operation, and good performance. For now we're focussing on simplifying our designs and manufacturing process, we'll figure out the rest afterwards."

"In that order?!" screamed another, sitting to the left. "Why you might as well replace the engine with a brick! If it's not able to drive you around, what even makes it a car?! Scrap your ramblings, first we make a car that works, with an engine and all, then we can figure out how to manufacture it."

"Pah," scoffed the engineer across the table, "and when is it in this story that you realize we're in the business of building cars, not mars rovers? If it doesn't get you from A to B faster than a bike nobody is going to give a hoot that it can run for three years without maintenance at the bottom of the Mariana Trench."

"Idiots!", the original butted back in. "If you make it simple first, then changing it to make it fast will be easy, and of course only simplicity begets correctness."

"You think the Mars Rover was simple?!─"

"Ha, because nuclear reactors are the paragon of simplicity─"

"CHERNOBYL IS EXACTLY THE POINT I'M MAKING HERE!─"

"Though wasn't it economic factors that lead to disuse of nuclear power? I hear solar is getting popular, we should really stick to my plan─"

"A car is not a solar panel, you bumble-headed fool─"

"You might as well be though─"

_Ahem,_ sounded the man at the head, drawing the room's attention, "I'm a little lost, so forgive the stupid question, but why have you not just considered... doing them together?"

Re: Simple, correct, fast: in that order

#317

Post by digsmahler »

Yes! So much covered in so few words. This ordering works shockingly well for constructing robust solutions to complex problems.

Keep it simple. Do what needs to be done now, leave off for later everything else. By the time later arrives, what the project needs will have gone in new and surprising directions. Refactoring complex and coupled code into logical discrete units pays off in multiples down the road. Removing that which is no longer needed is like giving your whole team extra time to breath and new room to think. For all the times that simplifying also solved the two other problems, we used that time for making more cool stuff.

Re: Simple, correct, fast: in that order

#318

Post by a1369209993 »

Const-me wrote: post #314
a1369209993 wrote: post #265

Earlier quoted context omitted.

> And yet, I don't see you or anyone else committing to buy ultra-simple non-pipelined non-OOO desktop/server CPUs. If you can find a CPU that has the same number of non-cache[1] transistors as a Intel/AMD chip, but spends them on a larger number of simple (and preferably independent/non-hyperthreaded) cores, rather than squandering them on speculative execution and ten thousand obscure model specific registers, I wo…

Intel makes them and you can buy them today, with up to 72 Atom CPU cores, e.g. (1) https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processo... Very niche products. For massively parallel number crunching, GPUs are much better in both performance/watt and performance/dollar. That Xeon Phi 7290 delivers up to 3.45TFlops, costs $3200, and consumes 245W. Compare with GeForce 1080Ti 10.6 TFlops, $700, same 250W. For…

https://www.intel.com/content/www/us/en/processors/xeon/xeon... shows at least six volumes of datasheets, and I still haven't found a instruction set refence. I have found https://www.intel.com/content/www/us/en/processors/xeon/xeon... (helpfully labeled "Datasheet, volume 2", rather than anything related to it's contents) which describes a subset of the aformentioned ten thousand random control registers. So no, Intel does not make [simple cores], it makes heaping piles of shit complete with malware ("Intel® Management Engine") buried at D22:F0 on a interal PCI bus.

It isn't simple, it's designed to be incorrect (and even the parts that are supposed to be correct aren't), and I'm not surprised it fails on fast as well.

Re: Simple, correct, fast: in that order

#319

Post by kungtotte »

sethrin wrote: post #280
kungtotte wrote: post #272

Earlier quoted context omitted.

I have a hard time seeing how profilers don't at least point you in the right direction. Or are we using different definitions of the word? They tell you how much time is spent in any given section of your program. If 80% of the time is spent in 20% of the code, it's usually a safe bet that's where you should start looking!

Profilers have limitations like anything else, and it's possible to be pointing the flashlight in the wrong place. I probably wouldn't include that as a list item. By the by, is there more than one kungtotte on the Internet? It took me a minute to think why that name was so familiar, but then I remembered watching a few hundred Beaglerush videos.

There must be more than one, because I've never heard of Beaglerush.

I've used this handle for a long time though (20 years or so), so it's all over the internet.

Re: Simple, correct, fast: in that order

#320

Post by Const-me »

a1369209993 wrote: post #318
Const-me wrote: post #314

Earlier quoted context omitted.

Intel makes them and you can buy them today, with up to 72 Atom CPU cores, e.g. (1) https://ark.intel.com/products/95830/Intel-Xeon-Phi-Processo... Very niche products. For massively parallel number crunching, GPUs are much better in both performance/watt and performance/dollar. That Xeon Phi 7290 delivers up to 3.45TFlops, costs $3200, and consumes 245W. Compare with GeForce 1080Ti 10.6 TFlops, $700, same 250W. For…

https://www.intel.com/content/www/us/en/processors/xeon/xeon... shows at least six volumes of datasheets, and I still haven't found a instruction set refence. I have found https://www.intel.com/content/www/us/en/processors/xeon/xeon... (helpfully labeled "Datasheet, volume 2", rather than anything related to it's contents) which describes a subset of the aformentioned ten thousand random control registers. So no, Int…

> I still haven't found a instruction set reference.

X86-64, SSE, AVX, AVX-512, AES-NI, etc. Their key selling point is software compatibility.

> Intel does not make [simple cores]

The cores are quite simple by today’s standards; otherwise Intel wouldn’t be able to pack 72 of them on a single chip. IME is unrelated to the cores, it’s a separate piece of silicon.

But if you don’t like the IME and don’t need backward compatibility with x86, maybe you’ll like this: https://www.qualcomm.com/products/qualcomm-centriq-2400-proc... But again, performance benefits of the architecture (48 simple cores) is questionable, GPUs are way faster for parallelizable number crunching, and you need single thread performance for almost everything else.

Post reply on HN