|
Should I take this as a precautionary note, since I re-watched The Matrix trilogy(*) this past weekend?
(*) Yeah, I know the first one is really the only good one. I just wasn't up to starting the next series in my usual winter binge-a-thon.
Software Zen: delete this;
|
|
|
|
|
Before I got bogged down with work I was working on two JSON projects - my own JSON parser and simdjson.
Both are very fast. simdjson is twice as fast as mine though. Mine keeps up with their nearest competitor.
Here's the thing though: I was running these on an old i5 with an HDD.
I had *no clue* how fast these modern machines were in comparison.
I mean, multiple GB/s even with my engine. It's unreal.
So what's the point of something like simdjson now? I was working on improving its performance even more, but why?
AMD was like, "nah - we got you. don't bother"
I feel thrilled and disappointed at the same time.
Real programmers use butterflies
|
|
|
|
|
Quote: So what's the point of something like simdjson now? I was working on improving its performance even more, but why?
Because you can't get no satisfaction.
"In testa che avete, Signor di Ceprano?"
-- Rigoletto
|
|
|
|
|
honey the codewitch wrote: I was working on improving its performance even more, but why?
Bragging rights? 
There is a point after which any further improvement is a waste of your time. If you use data only on your own platform, store it in the most convenient binary format. If you need to transfer it to another machine in a portable format, then JSON is an option. However, the time required for a network transfer even over a dedicated 1Gbps line will overwhelm the time required for parsing.
Freedom is the freedom to say that two plus two make four. If that is granted, all else follows.
-- 6079 Smith W.
|
|
|
|
|
Daniel Pfeffer wrote: There is a point after which any further improvement is a waste of your time.
Yeah. The issue was, it *wasn't* a waste of time on my older machine. I had no context for how much faster a modern machine was.
Daniel Pfeffer wrote: However, the time required for a network transfer even over a dedicated 1Gbps line will overwhelm the time required for parsing.
This is true, but for files the calculus is different, and part of what my library was designed to do was to parse through huge data dumps, typically in line delimited "JSON" format, so network times weren't really the bottleneck in that case. Even I/O wasn't entirely, even on my old machine, which surprised me.
Real programmers use butterflies
|
|
|
|
|
honey the codewitch wrote: So what's the point of something like simdjson now? I was working on improving its performance even more, but why?
AMD was like, "nah - we got you. don't bother"
'Cause not every situation has a new X86 CPU with bags of RAM and an SSD to throw at it. Sometimes you need good performance on minimal hardware with a small footprint.
Keep Calm and Carry On
|
|
|
|
|
The performance is already decent on my old machine, searching through JSON at almost 600MB/s
it also will run happily on an 8-bit CPU with <8kB of RAM, probably 4kB or less.
So I take your point, but on balance I think it's good. My rationale is this. If you need superfast JSON bulk parsing, you're going to buy a decent machine. If you're okay with an older machine, you're probably okay with 600MB/s of throughput. I think that's reasonable.
Real programmers use butterflies
|
|
|
|
|
honey the codewitch wrote: If you're okay with an older machine, you're probably okay with 600MB/s of throughput. I think that's reasonable.
I would be already happy with 100MB/s so... 
honey the codewitch wrote: If you need superfast JSON bulk parsing, you're going to buy a decent machine. If you're okay with an older machine, you're probably okay with 600MB/s of throughput. I think that's reasonable.
Yes it is. If the performance is a big step, then it is worthy. For just 1% or 2% or even a 5% faster... it is really to think if the time struggling with the improvement is worth or not
M.D.V. 
If something has a solution... Why do we have to worry about?. If it has no solution... For what reason do we have to worry about?
Help me to understand what I'm saying, and I'll explain it better to you
Rating helpful answers is nice, but saying thanks can be even nicer.
|
|
|
|
|
To be honest, this is why I don't develop on an uberPC - if it runs at an acceptable pace on my dev hardware, then it'll run well on any client kit.
"I have no idea what I did, but I'm taking full credit for it." - ThisOldTony
"Common sense is so rare these days, it should be classified as a super power" - Random T-shirt
AntiTwitter: @DalekDave is now a follower!
|
|
|
|
|
I feel the *exact* same way and have for as long as I've written code professionally.
However, my old machine was dying. I needed a new one, and buying a low to mid range machine is throwing good money after bad. Getting a midhigh or high end machine means it can remain viable for longer.
Although given my last machine died before the performance got unusable maybe I should take that into account.
Real programmers use butterflies
|
|
|
|
|
Quote: Although given my last machine died before the performance got unusable maybe I should take that into account. Smile |
My **current** personal machine (where I do all of my personal projects) is 10 years old (Gen-1 i7, 16GB RAM), and it's plenty fast enough for my current personal projects.
I'm looking at a new machine because now I need to do some Android development and there isn't an IDE for Android that doesn't require an UberMachine.
|
|
|
|
|
I think it's broadly true that there's not much point optimising most code given how fast machines are and how cheap space is. It's a long way from counting bytes in C++
|
|
|
|
|
I don't know. On my original machine my code ran at maybe 30MB/s before I got it up to almost 600MB/s.
That was worth it.
The other thing is the code will run on a machine with 4kB of RAM. I optimized for memory usage, primarily, not speed, so that this code could run on arduinos. Without optimization that wouldn't have been possible.
Certainly however, a lot of times it's not worth it.
Real programmers use butterflies
|
|
|
|
|
For a library, sure. For day to day code, not so much
|
|
|
|
|
Yeah, definitely. I can get behind that. My JSON reader is a library of course. I spend more time writing libraries than writing code that uses them. 
Real programmers use butterflies
|
|
|
|
|
Sounds fun
|
|
|
|
|
energy usage.
Performance is a clear functional requirement that must be met, energy usage is a secondary quality attribute (a.k.a. non-functional requirement) that has gained attention in the last 2 decades.
Computers, servers, internet, cloud, big data etc already use (fill in your favorite number)% of our total energy consumption and that number is very likely to rise. Energy efficient algorithms help to slow that growth.
Another area where energy efficiency is important is mobile devices. Some of you may recall apps on your phone that drained the battery. And IOT, Arduino etc, think devices that run on battery or solar.
As a rule of thumb, faster algorithms use less energy, so improvements in performance might still be valuable.
mental exercise
Another argument for improving performance of algorithms is the mental exercise which is fun for some of us.
|
|
|
|
|
rob tillaart wrote: Energy efficient algorithms help to slow that growth.
That's a very good point. Considering my JSON library is meant to run on devices like 8-bit Arduinos, ESP32s, hopefully android phones at least, and windows, linux and apple PCs it pays to make it scale down to the smaller devices while still being efficient
I do that, and my JSON parser is so small the average query takes far less than 4kB. Between the locality and the small RAM usage, I expect the power consumption to be relatively small on small machines.
On PCs the code uses memory mapped files for performance, which of course would increase energy use despite the speed increase due to paging and such, but it would be hard to tell without profiling for energy use - a kinda tricky thing if you've never done it before.
Real programmers use butterflies
|
|
|
|
|
Reminds me a bit of the ISP's hawking their wares. Lot's of TV adverts, emails, etc., to upgrade that 1Gb/s line yo always needed. Now, if you're a corporation with lots of users and lots of concurrent traffic - maybe.
But for us mere mortals? I noticed no difference when they upgraded me (oddly, to keep the price down with a new 1yr sprecial) to 300Mb/s . As that deal wound down I managed a new one, good until the end of 2022, for $10 less by dropping to 200Mb/s (and letting them charge my credit card automatically). Still no difference.
I mean, in the great reality of things - who can even send you data that fast? And, with net traffic, how would it get to you. And that's the parallel.
Scores of years ago, when I spent some time on Chemical Kinetics, that was always an importing thing to consider - or discover: "The Rate Determining Step" - and whatever is done around it will have no real effect. In layman's terms, it the intrinsic idea of a bottleneck.
| Ravings en masse^ |
|---|
| "The difference between genius and stupidity is that genius has its limits." - Albert Einstein | | "If you are searching for perfection in others, then you seek disappointment. If you seek perfection in yourself, then you will find failure." - Balboos HaGadol Mar 2010 |
|
|
|
|
|
Yeah. The bottlenecks in this case depend widely on what you're doing.
With my JSON lib:
If you're reading bulk data from a file on a modern PC, your bottleneck is both file i/o *and* CPU
If you're reading data from a file on an atmega2560 your bottleneck is SD i/o *and* CPU.
If you're reading data from a file on an ESP32 your bottleneck is primarily SD i/o
If you're reading data from a network on an atmega2560 your bottleneck is primarily network i/o *and* CPU
If you're reading data from a network on most any other device your bottleneck is the network.
I always try to know where my bottlenecks are. I always know where my towel is as well.
Real programmers use butterflies
|
|
|
|
|
honey the codewitch wrote: I always know where my towel is as we
I noticed, however, the conspicuous absence of your work with magnetic tape drives.
| Ravings en masse^ |
|---|
| "The difference between genius and stupidity is that genius has its limits." - Albert Einstein | | "If you are searching for perfection in others, then you seek disappointment. If you seek perfection in yourself, then you will find failure." - Balboos HaGadol Mar 2010 |
|
|
|
|
|
This is a GREAT question, and it is SOLUTION Dependent!
If you could make the GNU Compilers twice as fast, PLEASE DO. Every time!
Well, until the time savings is so tiny, nobody cares.
We have a phrase: "The right answer late... Is the wrong answer!"
To us, it means that timing is part of the design/interface.
A Ping response that is normally > network timeout is useless to the recipient, as they gave up!
A page that takes 1 second longer than a user will wait, is too slow.
Sometimes, it's about "feedback" (Working... Spinners, Changing Screens, see Travelocity, LOL)
Sometimes, it's simply too slow to feel useable. Or I can do it by hand.
When I can launch EXCEL faster than find my calculator... You can bet I will use excel.
Therefore, either FIND a situation that REQUIRES twice as fast JSON parsing (Apps/Databases/etc), or move on.
Personally, I won't optimize beyond "acceptable" without a business requirement. (NOTE: A business requirement can be indirect, like the impact having 1 million of these running every day at full load).
As an example, we have an update statement taking 2 seconds. It feels slow to me. It's processing almost a million records. Part of a bigger process that is taking too long by 30 minutes. Not in a loop. I think I'll pass. [There is an 8 hour solution that would remove this. I logged a note, not in this pass. Making that update twice as fast is meaningless in this context]
|
|
|
|
|
One of the challenges in developing my JSON processor was making it scale to 8-bit machines with 4kB of RAM reading an SD card and parsing through JSON at it a whopping 24kB a second. On my crappy old i5 that has since died I got 600MB a second throughput on the top end. On my new machine I haven't tried it yet, but with my NVMe and this processor i expect well over 1GB/s speeds - probably multiple GB/s
The issue here is there's two problems with the same solution.
One problem is getting a tiny device to be able to parse JSON that is effectively "too big" for it to use using traditional JSON parsing methods.
The other problem is getting more capable machines to process bulk data efficiently.
The solution is the same - an optimized streaming pull parser whose memory usage is *not* in any way tied to the size of the document, or even individual field lengths.
But since it's a library, and because of the disparate devices, plus tackling two different problems, what "optimized" is varies widely and is in the eye of the beholder I guess.
It would be really hard to come up with requirements around performance without a specific use case for it, if not impossible.
Real programmers use butterflies
|
|
|
|
|
honey the codewitch wrote: It would be really hard to come up with requirements around performance without a specific use case for it, if not impossible.
And neither having a specific use case is it guaranteed that you will come up with them...
M.D.V. 
If something has a solution... Why do we have to worry about?. If it has no solution... For what reason do we have to worry about?
Help me to understand what I'm saying, and I'll explain it better to you
Rating helpful answers is nice, but saying thanks can be even nicer.
|
|
|
|
|
Because if the JSON parsing is in an inner loop that's executed frequently you'll see serious improvements in your application.
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
