Why can't memory RAM be expanded externally?

[Dewdman42]

I didn’t say memory paging affects ssd performance.  I said ssd performance affects paging performance.  There have been some articles out there taking about m1 having better virtual memory performance compared to older macs and inferring that people are running happier with less memory than compared to older systems.  That is the primary question of this thread.
 

 Unfortunately I do not have any links to those articles for you now nor can I refer you to any related hard data without doing some research at this time which I don’t have time to do but I encourage you to if you are interested to know.

i am unaware of any actual compression that happens during virtual memory paging in the M1 macs but I would like to hear more about that.

Edited November 9, 2022 by Dewdman42

[analogika]

17 hours ago, Dewdman42 said:

I didn’t say memory paging affects ssd performance.  I said ssd performance affects paging performance.  There have been some articles out there taking about m1 having better virtual memory performance compared to older macs and inferring that people are running happier with less memory than compared to older systems.  That is the primary question of this thread.

Now I get why it feels like you're arguing with me, even though we're largely in agreement: This was never in question. Faster systems will be faster. Faster SSDs will be faster. Of course. 
All of that has been settled fact. 

 

Again: What the thread has actually been about the last few days is the claim that paging will cause the SSD to slow to the point where it cannot deliver audio in time. 
 

YOU didn't say that. Neither did I. MikeRobinson implied it. You're arguing with me about a point I never made. 

 

17 hours ago, Dewdman42 said:

Unfortunately I do not have any links to those articles for you now nor can I refer you to any related hard data without doing some research at this time which I don’t have time to do but I encourage you to if you are interested to know.

I'm well aware of reports that Apple Silicon seems better at memory management than previous Intel systems. 

I was asking specifically about the performance hit to SSDs when the system begins paging out. Given the massive parallelity of SSD access, my reasoning is that SSD speed is not going to bottleneck a DAW into failing before CPU overload does. 

17 hours ago, Dewdman42 said:

i am unaware of any actual compression that happens during virtual memory paging in the M1 macs but I would like to hear more about that.

Memory compression has been a cornerstone feature of macOS since Mavericks in 2013. The idea is to compress memory in RAM before paging out, trading off the CPU overhead involved with compression against avoiding the slow speeds of paging out to virtual memory. 

Details here: 
https://arstechnica.com/gadgets/2013/10/os-x-10-9/17/#compressed-memory

Edited November 10, 2022 by analogika

[MikeRobinson]

I would now like to clarify my previous statement.  Obviously, computer systems are now both "faster" and "more capacious" than they have ever been – or, that any of us probably ever dreamed that they could ever be.  Likewise, "SSD 'hard drives,'" since they actually consist of external memory-chips, are faster because there are no more "mechanical latencies."  Any "sector" of the available storage can be returned with no delays based on their location: the timing is now identical in all cases, and "non-mechanical."

But, all that having been said, "virtual storage" is still a carefully-crafted illusion.  If the requested piece of data is not actually in physical RAM, it must be retrieved: "paged in."  And, before that can happen, another piece of storage might have to be "paged out" to make room.  All of this takes time.  "Mere milliseconds," yes, but time nonetheless.

The operating system (MacOS) maintains this illusion on behalf of all programs, including Logic.  The programs are never explicitly aware that the trick is being played. To them, all of their "storage" is "real," and it has nothing to do with how much RAM is actually installed in the computer.

If Logic – for whatever reason – does not have access to any piece of data at the precise instant that it needs it, "System overload" is the result.  Obviously, this unfortunate outcome happens much less often than it used to.  (Yay!)

But, if it does occur, this is almost certain to be the reason: the CPU didn't "run out of horsepower." (As if, these(!) days, it ever possibly could ...)

It is for this reason that I always prefer "RAM size" over "CPU type" when making hardware purchase decisions.  I err on the side of caution.  I want my roads to have as many "lanes" as possible.

Edited November 10, 2022 by MikeRobinson

[Dewdman42]

2 hours ago, analogika said:

Details here: 
https://arstechnica.com/gadgets/2013/10/os-x-10-9/17/#compressed-memory

Thanks for that link I was not aware of the compression tech.  I wonder how I can monitor how much that is happening with my 5,1

seems to me the M1 must almost certainly be able to optimize memory compression by a significant amount also due to consolidated Memory and generally more cores to work with compared to older models.  In fact that might be even more significant then ssd improvements.

if I am restating what has been said in previous posts I apologize I am not interested at all in debating about whether virtual memeory slows down the ssd, I am interested in general ram requirements of Apple silicon compared to previous.

Edited November 10, 2022 by Dewdman42

[analogika]

3 hours ago, MikeRobinson said:

But, all that having been said, "virtual storage" is still a carefully-crafted illusion.  If the requested piece of data is not actually in physical RAM, it must be retrieved: "paged in."  And, before that can happen, another piece of storage might have to be "paged out" to make room.  All of this takes time.  "Mere milliseconds," yes, but time nonetheless.

It's actually MICROseconds, and it is completely irrelevant to Logic, because any disk latency is automatically compensated by Logic's playback buffer. What matters is BANDWIDTH, and that, as mentioned several times, is capable of several thousand tracks of the highest resolution currently used in any commercial recording system anywhere. 

Disk speed is simply not an issue. 

 

 

3 hours ago, MikeRobinson said:

The operating system (MacOS) maintains this illusion on behalf of all programs, including Logic.  The programs are never explicitly aware that the trick is being played. To them, all of their "storage" is "real," and it has nothing to do with how much RAM is actually installed in the computer.

If Logic – for whatever reason – does not have access to any piece of data at the precise instant that it needs it, "System overload" is the result.  Obviously, this unfortunate outcome happens much less often than it used to.  (Yay!)

But, if it does occur, this is almost certain to be the reason: the CPU didn't "run out of horsepower." (As if, these(!) days, it ever possibly could ...)

All of the above is completely incorrect, I'm afraid. 

Logic in fact has two DISTINCT error messages for either case: One is the "System Overload", and the other, believe it or not, is "Disk too slow". 

The very fact that you've apparently never seen a "Disk too slow" message is perfect proof that it simply is not an issue any more. This used to be super common in the days of ATA hard drives. 

System Overload is EXACTLY an overloaded CPU. And it is not necessarily the entire CPU, but a single core that overruns. 

This is really easy to achieve on any Intel Mac, btw.: Logic processes every channel strip on a single core of the CPU. Open the Performance meter. Create a new track with Alchemy, set it to highest quality, record a couple of chords, loop and press play.

Now add instances of Chromaverb to the channel strip inserts and watch ONE of the cores get loaded, while the others idle. 

Once that core hits 100%, Logic will fail. 

Logic works the same way on Apple Silicon, but from all accounts, it runs MUCH more efficiently there. 

 

 

3 hours ago, MikeRobinson said:

It is for this reason that I always prefer "RAM size" over "CPU type" when making hardware purchase decisions.  I err on the side of caution.  I want my roads to have as many "lanes" as possible.

Your image of "lanes" is the exact analogy for CPU cores. 

The priority is CPU cores, then RAM. More RAM is of no use if your CPU isn't equipped to process the content fast enough, with enough lanes. 

SSD, in turn, is only a consideration in terms of having enough storage. 

 

 

[Dewdman42]

 

On 11/10/2022 at 11:29 AM, MikeRobinson said:

If Logic – for whatever reason – does not have access to any piece of data at the precise instant that it needs it, "System overload" is the result.  Obviously, this unfortunate outcome happens much less often than it used to.  (Yay!)

This is where you are over-estimating the severity of the concern.  As I've tried to indicate earlier, Logic is not operating internally in a real time fashion.  There is not a "precise instant" that Logic needs certain data.  There is a length of time, determined by the size of your audio buffer, which LogicPro has time to compute all manner of things and fill the buffer with audio data that corresponds to a relatively long amount of musical time.  LogicPro can do whatever it needs to do at whatever pace it can do it, starting and stopping, filing up the buffer.  It can scream through some operations extremely fast and may take longer for others and there can be some amount of waiting around happening, this is all happening in the realm of microseconds.  If some memory needs to be decompressed, or paged in...there is some time to do it....there is no precise instant, when it will need that data instantly...there is a big wiggle room of time, which corresponds to the size of your audio buffer.  A larger buffer gives you more wiggle room.

Yes its true that if you have too much stuff paging out to SSD, you can reach a state of thrashing where it simply will not be able to do everything it needs to do in order to fill the buffer in time because its working over time...and so every little thing helps to make sure that will not the the case.  that includes BOTH making sure your CPU can keep up as well as it can be affected by memory speed, perhaps SSD speed if and when page files need to swap in, but also keep in mind, as already pointed out by others...there is a compression scheme which happens first..which is already helping to reduce the amount of paging...and also MacOS is smart enough to compress or page out memory which has not been accessed in a while.  The point is, a little bit of paging may not kill you as much as you think.  

On 11/10/2022 at 11:29 AM, MikeRobinson said:

But, if it does occur, this is almost certain to be the reason: the CPU didn't "run out of horsepower." (As if, these(!) days, it ever possibly could ...)

 

It very well could be CPU.  Or it could be other things.  The CPU has to wait on other things and sometimes the other things are waiting on the CPU...so its all relevant.

 

On 11/10/2022 at 11:29 AM, MikeRobinson said:

It is for this reason that I always prefer "RAM size" over "CPU type" when making hardware purchase decisions.  I err on the side of caution.  I want my roads to have as many "lanes" as possible.

I'm not sure I agree with that assessment of preferring RAM size over CPU type.  It all matters and if price is no object then get as much RAM as you can get of course, but anyway with Apple Silicon our options are going to be a lot less.  And quite a bit more expensive to get more RAM then you really need.  I think the memory management in Apple Silicon is much better than older systems, both because of the consolidated chip design as well as faster bus speeds in both the memory and SSD.  All of that equates to systems that can do quite a bit more work in less time with less memory, than has previously been the case in the past.  So in the past you might have thought 64gb was mandatory for daw work but now might be fine with 32gb.  I know of people running MaxStudio with 64gb and have been mixing large orchestral mixing sessions with hundreds of tracks of plugins and absolutely no complaints...all running like a charm.  If you are not running hundreds of orch sample tracks..you probably can work just fine with 32gb.  The various architectures that have all been discussed here are all reasons that Apple Silicon machines can run very effectively with less memory than previous Mac models.

Eventually, you would hit some upper limit on these too, where it just can't keep up with some massive load, even with 64gb of RAM.  I think the 128gb model is entirely overkil for 99% of us thoughl!  I think its likely that even the 64gb model is overkill for a lot of us, but me personally I will probably get the 64gb model because I prefer to know its not having to page anything at any time or run into any limitation due to not enough RAM...especially since we can't install more ram after the fact.   So I feel you on that.  But at the same time, pay attention to what other users with MacStudio's are already doing with 32gb and 64gb and realize that the RAM requirements will not be as much as previous Mac models.

My 12 year old 5,1 MacPro does quite well with only 96gb of ram...  I think the MacStudio with 64gb probably does at least as well or better then that in terms of memory usage, just due to the factors we have discussed.

But hey, if you have pockets of cash burning holes in your pockets....go for the Ultra 128 and never lose any sleep about it.  

But adding non-consolidated memory to the model is also not the right solution.  All memory access to RAM outside of the CPU will be much slower then when its directly in the CPU.  If you want Apple to make M2 or M3 with more memory...maybe they will, but I don't think so honestly.  The very vast majority of users out there don't need more RAM than that.  And that includes the vast majority of audio production users also.

 

 

Edited November 11, 2022 by Dewdman42

[olavsu1]

It's possible only client-server  systems. 

in client side they don't have ram and hdd's.  only boot rom  and IO.

all  app's  and  ram, hhd's, gpu/cpu units/nodes  are in server and it is large thing.

//I have seen one external 1TB memory module up close myself, it was 6U and so heavy that it had to be lifted by four men.

[analogika]

14 hours ago, Dewdman42 said:

Yes its true that if you have too much stuff paging out to SSD, you can reach a state of thrashing where it simply will not be able to do everything it needs to do in order to fill the buffer in time because its working over time...

This keeps getting repeated, and I have yet to see any evidence that it is true.
Thrashing — down to the literal word itself — is a relict of the age of mechanical hard drives. 
 

It it really even a theoretical concern at all on a Mac with 20-Gigabit/sec throughput? 

Yes, paging will slow the computer’s performance. Of course. 
 

But the repeated claim that paging can drop DISK performance of an internal SSD to the point where it’s too slow for audio performance on a buffered system really needs to be supported by some sort of evidence, IMO. 
 

 

[Dewdman42]

I wasn’t saying thet.   You keep repeating this thing about something affecting ssd performance and I don’t know why you keep saying that, I certainly am not implying that and have not said so even once.

If you theoretically use too much memory in a patellel fashion in a system with not enough ram EVENTUALLY the system will thrash, that is not a relic term! I’m just saying it’s not going to happen “the instant” the first memory paging operation occurs, there is a lot of wiggle and breathing room.  Thrashing is the outer extreme that most of us have avoided by having enough ram in our systems to do everything we need to do.

this has sure become and unfriendly thread.  I’m out.

Edited November 12, 2022 by Dewdman42

[analogika]

Sorry about seeming "unfriendly". 

I'm actually just trying to find real information, so I'm asking people to substantiate their claims — several posts in this thread are advising people to spend hundreds of dollars applying criteria that simply do not seem to describe real problems — in the context of audio production. 
 

 

[MikeRobinson]

I respectfully yield ... You are much-more correct ...