Lecture 009

Memory Abstraction

Memory in Practice

S-RAM(MB): expensive, big, fast, used in cache D-RAM(GB): cheap, small, slow, main memories, frame buffers

• Synchronous D-RAM: switch to use clock signal

• Double Data-rate ynchronous D-RAM (DDR SDRAM): double edge clocks send 2 bits per cycle per pin (DDR = 2 bits, DDR2 = 4 bits, DDR3 = 8 bits, DDR4 = 16 bits) - standard in servers and laptop (i7 support DDR3 and DDR4)

Operations in DRAM

Reading

• 1-a: Row access strobe (RAS) select row 2

• 1-b: Row 2 copied from DRAM to row buffer

• 

• 2-a: Column access strobe (CAS) select column 1

• 2-b: Supercell of size (2, 1) copied from buffer eventually to CPU

• 2-c: Row buffer copied to DRAM to provide refresh

• 

• 3: Continue copy to buffer and do its refresh as normal (if you access the next thing that will refresh, code can run faster because it does not disrupt normal refresh schedule)

64MB memory module has 8x 8M DRAMs to provide 64 bits data in parallel

• 

Locality

• Memory Hierarchy: we can use locality to fill the gap by creating device that fill the gap

Principle of Locality: programs tend to use data and instructions with address near or equal to those they have used recently

• Temporal locality: recently referenced items are likely to be referenced again in the near future

• Spatial locality: items with nearby address tend to be referenced close together in time

• 

• Ideally: memory hierarchy creates a large pool of storage that costs as mush as the cheap storage near the bottom, but that serves data to programs at the rate of the fast storage near the top

Cache

Cache (L1, L2, L3): usually made of S-RAM, usually in CPU core, to provide fast cache from D-RAM.

• 

• 

Cache Policy

...

Cache Misses

Cold (compulsory) miss: when starts empty, first reference to the block (nothing in the cache right now)

Capacity miss: when working set is larger than the cache (no enough cache space to store all needed data for a program, cache is full and your data is not here)

Conflict miss: when multiple objects all map to the same level k block (hash collision: block i at Lk+1 must be placed in block i mod 4 at Lk, referencing blocks 0, 8, 0, 8,... will miss every time)

Storage

Disk

Capacity: maximum number of bits that can be stored.

• Recording density (bits/in): number of bits that can be squeezed into a 1 inch segment of a track.

• Track density (tracks/in): number of tracks that can be squeezed into a 1 inch radial segment.

• Areal density (bits/in2): product of recording and track density.

• Seek Time (moving head): 3-9 ms

• Rotational Latency (rotating disk): 4 ms

• transfer time = 0.02 ms

• average time = 13 ms (first bit is expensive, the rest are free)

Reading a Disk to Memory:

1. CPU send command to disk controller
2. Disk memory transfer to Main memory
3. Disk controller send interrupt notification to CPU

Writing to a Disk from Memory:

1. CPU tell disk a specific memory location
2. Disk read memory out to Disk

Non-volatile Memory

Non-volatile: retain value even if powered off

• ROM stores: BIOS, controllers for disks, network cards, graphic accelerators, security subsystems

• SSD

• Disk Cache (for accumulating read/write commands to make disk faster)

Types of Non-volatile

• Read-only Memory (ROM): programmed during production, usually launching OS

• Electrically erasable PROM (EEPROM): ...

• Flash Memory: EEPROMs, with partial (block-level) erase capability (only 100,000 erasing)

• 3D XPoint (Intel Optane) & emerging NVMs (New materials)

Solid State Disk (SSD)

• read / write in unit of pages (512 KB ~ 4 KB)

• pages can be written when its block (32 ~ 128 pages) has been erased (a long time like 1ms)

• advantage: faster, less power, no head crash

• disadvantage: wear out faster (mitigated by "wear level logic" in flash translation layer)