I don’t think this was Steve’s best video. In my opinion he mixed up three important concepts that have different motivations and are easier to understand separately: relocation, virtual memory and segmentation. Early computers ran a single job at a time so the programmer could decide to start his program at any arbitrary address and place the data in memory as he wanted, there was no interference from other programs. The next step was multi programming where several programs were in memory at the same time and the processor time-sliced between them. Initially this was achieved by using a ‘relocating loader’ that adjusted any internal (relative) address references to the actual addresses in memory. Then hardware was devised to help - a Base and Limit register set that allowed each program to see its address space starting at zero and continuing up to the limit. This had the dual benefit of performing the relocation using hardware that just added the base address to each memory reference (and checking it didn’t exceed the limit) and protecting other programs (and the operating system) from getting overwritten by accidents or bugs. The Base/Limit approach continued into the 1970s (think ICL1900, IBM360) but had the disadvantage that memory needed to be allocated in one contiguous block for each program, which leads to inefficient use over time, unused gaps etc. And this at a time when memory prices were the key constraint in building large computers. To increase efficiency the team at Manchester University designed Atlas to have demand paged virtual memory, so that the need for large contiguous blocks of memory could be overcome. Finally the next generation of computers using virtual memory, eg Multics figured that you could also Segment virtual memory so as to have read only pages shared by several programs, like library routines, and other uses for shared segments were found, for inter program communications. Hope this helps.

Another benefit of virtual-memory is memory-isolation. It prevents programs from messing with memory of other programs, which increases stability and increases security (it's not 100%, but certainly better than letting all programs access shared physical memory).

Someone take away his pastel coloured markers and give him a black pen - then we'd be able to see his diagrams!

6:40 There's a real benefit for multiple mappings and it's when running virtual machines: the virtualised OS does one mapping for its programs and the host OS does another to the physical RAM.

I've had the question in the title of this video for numerous years. Thank you Computerphile.

The presenter is getting very good at asking pertinent questions. If we ever have an AI working on the meaning of life the universe and everything he could come in handy.

The lighting + fluorescent marker was a bold choice. But I didn't dislike it as much as others seemed to have. I think it was fun and interesting, in small doses. (I certainly wouldn't want it to be the norm though.)

I liked that the talk mixes different concepts and approaches on RAM management - all because it has been developed on top of one another and tries to solve one common problem, in steps.

Windows 10 and 11 use this technique as "fast startup", too. So, when you shut down the PC, it writes the volatile memory to the non-volatile storage. That way, your boot time is quicker. Rather than going through your entire bootloader sequence, it can just throw the written storage DATA back into RAM. Kind of like a hybrid of Hibernate mode and Sleep mode. This is why Hibernate mode is now hidden by default in Win10/Win11.

About 30 years ago I wrote an incremental linker for Unix (mostly for Sun, but it got ported at least a couple of times -- for all I know, it's still running someplace), and this was around the time that a lot of low-level I/O operations were starting to be written using mmap. Some colleagues suggested at the time I should forsake read() and write() and just use mmap for everything, but my experiments showed it actually made the linker slower to use that for everything, despite it being faster to use it for some things, such as incrementally updating the debugging information in the output file. I've occasionally wondered if making basic I/O depend on mmap ever happened, and if so, how they solved the performance issue I saw. But I don't program at the level anymore, so I don't know how things work these days, and I'm at the mercy of the people writing O/S stuff, instead of being one of the people inflicting my decisions on others.

Starting the weekend well!

FYI, one of the most popular home computers of the 1980s, the Commodore 64, was equipped with 64 KiB of RAM. Its CPU, the MOS 6510, divided memory into 256-byte pages. That's 8 bits of address space per page. Makes a lot of sense, considering it was an 8-bit CPU (other than addresses, of course). It even had a special "zero-page" addressing mode for some instructions, that reduced the instruction size by a byte, and often saved a cycle too, in exchange for being restricted to the lower 256 bytes. As you might imagine, that zero page was crowded with important system variables. Well, the C64 also had no virtual memory. But then, it barely had an operating system, either, so...

I think it was a Sun OS implementation where we put a lot of Ram in this particular system and the OS insisted that you had to have a swap partition about 1.5x the real Ram. It meant we had to buy an extra ultra scsi disk just for a swap partition we didn’t even want. Learning experience

There are actually microcontrollers with 4 bytes of RAM. They use flash as program and static data memory, and have 1 kilobyte of it; RAM is used only for changing data, e.g. in the hello world you'd need 1 byte of RAM to loop over the data. They are useful in smallest of embedded applications: battery monitors, pushbutton inpts, watchdog timers, simple regulators, infrared remotes and so on.

finally i can leave hundreds of tabs open on my ipad exactly like my pc

Who makes these thumbnails? That person needs a raise.

Virtual memory makes me think of "downloading more ram".

Great video!!!

Guuyyyys, give the man a high contrast marker so we can see what he draws.

Virtual memory is also important for making 32 bit programs work seamlessly on 64 bit systems. Each program can still only target 4GB, but by translating the addresses to 64 bit, they can be spread out in physical memory and avoid overlapping memory spaces.

I actually made a video on this lol but without all the jargon. Lot's of stop gap animation as well. But listening to a master of his craft is awesome.

11:07 Love how he just casually explained the concept of address mirroring. This is actually a trick the Game Boy Advance uses: A pointer to the IRQ handler is supposed to go at 0x03007FFC but an actual interrupt loads the address stored at 0x03FFFFFC which is much easier for the ARM to load due to its use of ompressed operands. Thanks to mirroring it works out to be the same value regardless.

Ben Eater’s 8-bit computer, 16 bytes, hello world works.

This video would have been *really* useful to me a week ago when I was having my OS finals

So it's kinda like a hashtable but for hardware?

17:24 Linux also supports swap files. In fact you can have many swap partitions and files at the same time and even add and remove them on the fly.

"of course 16 byte CPU wouldn't be much use for anything" -- the guy obviously hasn't heard of Ben Eater

The good thing about swap files on linux is that you can add them on the fly if you need them and if your computer is still responsive. Swap files and partitions can also exist simultaniously for whatever reason there might be for such a setup. But I have a question about swap files, is there a way to read what data is stored in it? I mean, could a swap file be a potential security risk?

In a way, memory chips do this all the time, since there is the address map, the logical map, and the physical map in the chip. The address map is what the CPU see as the memory locations, e.g. 0x0023-0x0025. The logical map is how the addresses are in the chip the address sequence might be 0x0023, 0x0025, 0x0024 for instance. and the physical map is how the actual memory cells are laid out on the chip, usually bit grouped instead of byte grouped, all the first bits are grouped, then the second bits are grouped, etc. Again, the order of the bits might be different on the chip as well, with the physical locations of the third bits of the byte are next to the first bits instead of the second. I'll admit I'm a bit out of date w/regards to memory layout since its been 21 years since I worked on testing the chips, so it might not be the same anymore.

When I first learned about virtual memory (in the context of 386 protected mode) all the pagetable stuff seemed like a bunch of convoluted nonsense. Now, I see how elegant it is, page faults transferring control to the OS to load pages of memory from disk into RAM, and the program continues on its way none the wiser.

I thought paging files/partitions were mostly phased out due to SSD wear leveling concerns? I know this was a big issue with running windows XP/Vista on SSDs (and whether or not to have a swap partition on linux systems).

01:50 I think you just named a semiconductor company. I can see it now; *_Luminous Computers_* in some warehouse district, with a giant sign in Neon (real or simulated).

I am more shocked iOS did not support swap/virtual memory/page before now.

I'd like to know more about a hypervisor running multiple operating systems with identical read only blocks. Does it dedupe and combine blocks across VM guests during idle time? And does VTd speed up hypervisor memory lookups?

Can I finally download more RAM?

Segmented addressing. Never mention that again! :)

Have you ever considered a series for GPUs? I think it would be great, since there are barely any reliable and modern information about how GPUs actually work nowadays.

I don't know if I got correctly what the professor said about the multilevel translation but modern systems do use it. Each process commonly uses only a little of the available memory and its "pages" are adjacent which means that its translation table is mostly empty. This would waste memory to store the empty cells of the table. To counteract this, translation can be made multilevel, so that each table points to another, which exists only if the memory it points to is actually used.

1:35 You would likely need to salvage part of the initial program space, overwriting it with the output. Donald Knuts book had a challenge, writing the shortest program that fills the entire memory (including the program itself) with a single value. This would be similar, except no perfect overwriting of every memory cell is necessary.

what if there was some important data like passwords, keys can u force it to allocate on disk and then pull the plug and search on disk for this important data?

Can you do a video on Microsoft patches and how they protect the computers from vulnerabilities

Why is this being drawn out on this kind of paper and highlighter? Was there no pen available? This is so hard to read

This seems close to how cartridge game console like the NES and SNES worked. The cartridge basically is part of the memory and can be addressed immediately. So nothing needed to be loaded into memory; the game cartridge is the memory.

They've boosted the saturation of orange to make the pen look radioactive, but the paper is still unreadable with insufficient contrast in brightness. Isn't this what was used to be called a memory mapped file? Programs dealing media streams wouldn't attempt to load the whole file into memory, unless the program is a dumb as Windows Sound Recorder. They'd read it from disk and use a structure like Photoshop's scratch disk for edits.

If you ever have run a program with AddressSanitizer enabled, then you get funny amounts of virtual memory used in htop. But physical memory usage is still normal.

That's a pretty trippy marker! I want one! 😛😉

Orange highlighter might not have been the best choice

Is there also wear levelling for RAM? And does that use a second virtual memory lookup?

To answer this question, I paraphrase Seymour Cray: "Virtual memory is memory that you don't have." - what he actually said was: "Memory is like an orgasm. It's a lot better if you don't have to fake it."

Wouldn't a similar thing happen when you attach a second RAM, so its starter address has to be appended to machine's total available?

How much delay does virtual memory introduce versus accessing memory directly?

Isn't dual abstraction of address space a big thing in security. I could have sworn the second layer was undersubscribed with blocks acting basically as honey pots. When the first layer tried to manually request the location in a rude way, think off by 1 C index requests, a security exception would be raised and the OS would know that there was a bad actor thread running. Please see buffer overflow vulnerability mitigation.

Didn’t IBM do an early version of this in the 1960’s.

Is this available on the M1 chip?

The implementation on iPad must be something more than a decades old computer technology? I can't get my head around this.

I've done plenty of things in assembly in embedded applications using 16 bytes or less. So it certainly isn't useless.

Would this somehow protect against attacks in memory tho? Meaning if someone is attacking the memory to try and get some data from there, and there is a virtual memory in place, would that somehow slow them down or prevent them from getting data out? Since they would be seeing one address but in fact its in another address. Or would it not matter?

Advertising your computer having virtual memory as a selling point in 2022 is like treating a car having wheels as a selling point.

How many layers of abstraction are possible before nobody in the world actually understands how a computer or application works?

Now I know why old computers gain so much performance from SSD upgrade due to the bit of memory being copied in a storage as a form of a translated pages (swap/ swap compressed). Accessing the HDD going to be insanely slow if you're RAM limited thus replacing it with SSD going to make a significant improvement on getting the CPU fully loaded with process. Especially on 32-bit machine with 4GB of memory. Of course we are still limited by the maximum bus speed at the end of the day even with, say, SATA3 ssd.

Nice

I think that safety is important and one should use double-protection technologies, such as Oreol Staking!

wow cool vid

Oh my! That pen...

Steve has the coolest shirt on

Virtual memory is backup RAM for when you play modern videogames like Warzone on an 8GB PC at the cost of fps.

You could probably do Hello World with 16 bytes because there are only 7 different characters. The rest of the bytes would be for mapping the duplicates

A perfectly excellent description of paging - which has been around for zonks. Sorry is there something "new", couldn't see that bit :S

Every once in a while Apple or Google announces a feature in one of their devices and my only reaction is "HOW did you NOT already have that?!"

Highlighter not the best pen 😀

VERY 🤔 about Samsung's "RAM Plus" feature. Feel like their offering less RAM then normal then replacing it with this potential hoodwink. My thought is wether it's innovative or a just a gimmicky thing.

Isn’t that essentially what mmap in c is doing?

Do this damage ssd

Download more ram?

Couldn't follow along with the yellow marker on white paper, very hard to read

VMware virtualisation / MS hypervisor Mapping changes all the time cause DRS moves workloads around and re-registeres virtual memory continuously

Sharpie please

Black marker pls

4K is the largest amount you can do 16 bit CRC with.

Hello world in ascii might work but ut8 🤷🏽‍♂️

This video screams "computerphile: late night". I love it, but it looks like you pulled Stave from a party after a few drinks and decided to make the video right then

The opposite of this is a RAM drive but with 12 terabyte hard drives being economical they don't get much use anymore

Fyi, orange highlighter is not the best choice for showing info on paper.

Memory Mapping ? is that the same?

do a video about the google AI being self-aware, please!

It's very confusing that the OS will somehow manage the logic of Virtual memory as OS is also running in the CPU and possibly using the data from RAM itself.

I am colorblind. Cannot see any of these diagrams.

Highlighters are supposed to be the most visible pen

Reminds me of the old 'Download extra RAM' popups. 😆

Is it just me or does apple sound really outdated? Like in this video virtual memory is mentioned as a new thing on iPads, and in another video a modern looking Mac was mentioned as having a 2 core cpu. Both of these things have been popular since the early 2000s.

2:29 Am I high? (The marker)

Who diagrams with a highlighter???

can we take a look at PassKey announced by Apple and maybe Google, Windows soon. hardware security primitives what are they?

The lion kingdom has always turned off swap space for solid state storage, out of fear of using up the write cycles. Not sure if they're doing it for planned obsolescence or because people have enough money to just not care about write cycles.

Drawing in orange highlighter was a bad move. You’d think they’d realize that the camera wasn’t going to pick it up well.

10PRINT"HELLO WORLD" on my BBC Micro takes 20 bytes (The difference between TOP and PAGE).

The opposite of RAM disc.

Genius

Virtual memory is the best

LFG if you don’t work with Oreol Staking provider yet