Introduction

• Why this lecture?
  • To understand what is happening in the “hardware” part of the execution stack
  • To write programs that are efficient on modern machines

Moore’s Law

• 1965 - 2005

  • Moore’s Law (1965): the number of transistors in microprocessors doubles every two years
  • The fineness of the processor engraving decreases
  • The clock frequency increases

  ⟹ Increased processor performance

• Since 2005

  • The fineness of engraving continues to decrease (but less quickly)
  • The clock frequency no longer increases due to heat dissipation
  • Heat dissipation depends on the frequency, and the number of transistors
  • Multiple computing units per processor

Evolution of processors performance

Source

Sequential processor

• An instruction requires N steps
  • Fetch: load instruction from memory
  • Decode: identify the instruction
  • Execute: execution of the instruction
  • Writeback: storage of the result
• Each step is processed by a processor circuit
• Most circuits are not used at every stage → One instruction is executed every N cycles

Instruction pipeline

At each stage, several circuits are used

→ One instruction is executed at each cycle

Micro architecture of a pipeline

• Each stage of the pipeline is implemented by a set of logic gates
• Execute step: one subcircuit per type of operation (functional unit)

Superscalar processors

• Use of different functional units simultaneously

⟹ several instructions executed simultaneously!

• Require to load and decode several instructions simultaneously

Superscalar processors throughput

Dependence between instructions

Limitations of the superscalar:

• There should be no dependency between statements executed simultaneously.

• Example of non-parallelizable instructions

a = b * c;
d = a + 1;

• Degree of parallelism of the instructions: Instruction Level Parallelism (ILP)
• Instructions executed in parallel must use different functional units

Branching

• How to fill the pipeline when the instructions contain conditional jumps?

    cmp a, 7    ; a > 7 ?
    ble L1
    mov c, b    ; b = c
    br L2
L1: mov d, b    ; b = d
L2: ...

• In case of a bad choice: the pipeline must be “emptied”

⟹ waste of time

Branch prediction

• The processor implements a prediction algorithm
• General idea:
  • For each conditional jump, store the previous results

0x12 loop:
         ...
0x50     inc eax
0x54     cmpl eax, 10000
0x5A     jl loop
0x5C end_loop:
         ...

Vector instructions

• Many applications run in Data Parallelism mode
• Single Instruction, Multiple Data (SIMD): the same operation applied to a set of data

for(i=0; i<size; i++) {
   C[i] = A[i] * B[i];
}

• Example: image processing, scientific computing

• Using vector instructions (MMX, SSE, AVX, …)

• Instructions specific to a processor type

• Process the same operation on multiple data at once

for(i=0; i<size; i+= 8) {
   *pC = _mm_mul_ps(*pA, *pB);
   pA++; pB++; pC++;
}

Parallel Processing

Hyperthreading / SMT

• Problem with superscalar / vector processors:

  • The application must have enough parallelism to exploit
  • Other applications may be waiting for the CPU

• Simultaneous Multi-Threading (SMT, or Hyperthreading)

  • Modify a superscalar processor to run multiple threads
  • Duplicate some circuits
  • Share certain circuits (eg FPU) between processing units

Multi-core processors

• Limited scalability of SMT

• dispatcher is shared

• FPU is shared

→ Duplicate all the circuits

Symmetric Multi-Processing (SMP)

• Multiple processors sockets on a motherboard
• The processors share the system bus
• Processors share memory
• Scalability problem: contention when accessing the bus

NUMA architectures

• NUMA nodes connected by a fast network
• Memory consistency between processors
• Privileged access to the local
• Access possible (with an additional cost) to memory banks located on other nodes

→ Non-Uniform Memory Architecture

Memory hierarchy

Memory wall

• Until 2005: increase in CPU performance: 55 % / year
• Since 2005: increase in the number of cores per processor
• Increased memory performance: 10 % / year
• The memory accesses which are now expensive: Memory Wall
• Mechanisms are needed to improve memory performance

Cache memory

• Memory access (RAM) are very expensive (approx. 60 ns - approx. 180 cycles)
• To speed up memory access, let’s use a fast cache memory:
  • L1 cache: very small capacity (typically: 64 KiB), very fast (approx. 4 cycles)
  • L2 cache: small capacity (typical: 256 KiB), fast (approx. 10 cycles)
  • L3 cache: large capacity (typically: between 4 MiB and 30 MiB), slow (approx. 40 cycles)
• Very expensive hard disk access (SWAP): approx. 40 ms (150 μs on an SSD disk)

Memory Management Unit (MMU)

Fully-associative caches

Direct-mapped caches

Set-associative caches

Cache consistency

• What if 2 threads access the same cache line?

• Concurrent read: replication in local caches

• Concurrent write: need to invalidate data in other caches

• Cache snooping: the cache sends a message that invalidates the others caches