Using Slidy

The Memory Hierarchy: How Computer Memory Is Organized

The fastest kind of memory sould actually be below the cache in the above diagram and consists of the registers of the CPU. Usually, one has around 16, 64bit registers. These registers can be accessed at close to the clock speed of the CPU, so on modern machines on the order of a nanosecond (`10^{-9}S`) or slightly faster.

The cache is part of the computer's microprocessor.
It is capable of holding both data and instructions.
The cache though only has limited memory.
Sometime cache changes are not immediately reflected in main memory.
A cache might have up to four levels: on-board L1 cache on the microprocessor, and levels L2, L3, and L4 which might be either on board on a different chip.
Cache memory takes on the order of nanosecond to tens of nanoseconds to access.
As an example, the Intel Core i7-8700K Coffee Lake (2017) and Core i7-10710U Comet Lake (2019) CPUs use a 14nm process, have 6 cores, 384KB of L1 cache, 1.5MB of L2 cache, and 2 MB of L3 cache. Each core has its own L1 cache, pairs of cores share L2 caches, and the L3 cache is shared.
Main memory look up is an order of magnitude slower than cache memory look up.

Nowadays, new PCs have between 8GB to 64GB of main memory.
A bit of memory can store a single binary digit. I.e., can hold either a 0 or a 1. A bit of of memory is abbreviated by a lowercase letter b.
A byte can hold 8 bits. A byte of of memory is abbreviated by a uppercase letter B.
For larger amounts of memory, either metric or a power of 2 notation is used.
Metric prefixes like kilo-, mega-, giga-, etc., we write with their standard initials k, M, G, etc.
Corresponding to each of these metrics, we have a corresponding base 2 prefix: kibi- (`2^{10}`), megi- (`2^{20}`), gigi- (`2^{30}`), etc., which we abbreviate as Ki, Mi, Gi, etc.

Putting the above together, we can make the following memory table:

Moore's law says the number of transisters in an integrated circuit that can be packed into the size package doubles roughly every two years.
Physics will eventually prevent this law from continuing indefinitely (we're probably close to its end), although it has held for the last 50 years.

Death Master File (SSNs of people who have died in US since start of social security)	~1.5Gb
UTZoo Archive compressed posts of Usenet from 1981 to 1991	~3Gb
Mozilla Addons circa 2008	4Gb
Compressed English Wikipedia Page Dump (not indexed)	15Gb
Open Street Map circa 2013	3Tb
All Wikipedia as stored in DB Excluding Blobs circa 2012	5Tb
Indexed 1 billion page web crawl on yioop.com circa 2014	24Tb
Text Portion of Library of Congress circa 2010	20Tb
All Wikipedia as stored in DB with Blobs circa 2012	27Tb

Programs make use of a virtual addressing scheme for storing data during program execution.
The goal is to try to fit as much data in memory as can at one time. If a program needs to look up some data that is not in memory then the CPU goes to disk to either read or store it.
For older 32 bit processors, one word can store up to `2^{32}` different numbers.
For modern 64 bit processors, one word can store up to `2^{64}` different numbers.
This word can be used as an address for the location of a memory word.
Somewhat counter-intuitively, software in use often changes much more slowly than hardware, so many operating systems, or software written for them, still make legacy 32 bit assumptions, so we'll use it as an example. Also, the numbers are smaller.
For a traditional 32 bit machine, a process would pretend it had 4GiB (the number of bytes addressable with a single word) of main memory (even if one had less) and would swap blocks of memory into and out of main memory to disk corresponding to data that was immediately needed.
Swapping between memory and is disk done using page sized blocks (4KiB to 56KiB).

In our memory hierarchy, we have indicated some layers are volatile and others are non-volatile.
Roughly, volatile memory is memory that one can expect to not-remain correct when the power goes off on the computer.
On the other hand, non-volatile memory is expected to persist even after a sustained power outage.
For most DBMS operations, we view the completion of the operation as occurring when the operation's effects have been committed to non-volatile memory (secondary storage).