Spaced Repetition

What is a parallel computer?

A system that runs multiple processing elements at the same time to solve one computational problem, using techniques that make it more performant.

What is the difference between concurrency and parallelism?

Concurrency: one core runs task 1, then task 2, then task 1 again, until both finish. Parallelism is that plus hardware: task 1 on core 1 while core 2 runs task 2. One cook switching between two pans is concurrency. Two cooks and two pans at the same moment is parallelism.

What is a machine cycle?

One clock tick inside the processor.

What does clock rate measure?

How many ticks per second. A higher rate lets one core do more work per second, until the chip overheats.

What is throughput, and what is FLOPS?

Throughput is how much finished work you get per second. FLOPS is how many decimal operations finish per second.

How does a parallel system differ from a distributed one?

Parallel systems chase performance: shared memory, fast links, one shared job. A weather model on a supercomputer is this. Distributed systems chase scale, sharing, and staying up: many machines, own memory, slower network, no single clock. A bank is this.

How do parallel systems and distributed systems differ in terms of memory, clock synchronization, and communication links?

Parallel: tightly coupled shared memory and a global clock, links on the order of Tbps. Distributed: distributed memory, no global clock so sync algorithms are needed, links on the order of Gbps.

How do serial and parallel processing differ?

Serial: instructions wait and run one after another. Parallel: instructions run at the same time.

What are the four levels of parallelism?

Job: separate scheduled jobs. Program: parts and threads. Instruction: the compiler and processor overlap instructions. Bit: vector operations and arithmetic on bits.

What do IPC and IPS measure?

IPC is instructions per machine cycle. IPS is IPC times the clock rate, so instructions per second.

What does Moore's law say?

The number of transistors in microchips doubles about every 2 years.

Why did chips add cores instead of raising the clock?

Raising the clock overheated the chip and used more power. Extra transistors went into more cores.

What does a circuit breaker wrap and what does it count?

It wraps a protected call like an external supplier or API call and counts failures.

What happens when the circuit is closed?

The call reaches the supplier, where a failure increases the failure count and a success resets it to 0, opening when the threshold is reached.

What happens when the circuit is open?

The call is not made and the client gets an error immediately.

What happens during the half-open state after a reset timeout?

A trial call is made, where success marks the state as closed and failure makes it open until the reset timeout finishes.

How can a hanging remote call cause thread exhaustion and cascading failures?

Many callers waiting on it take new threads and exhaust the thread pool, which can also trigger the breaker.

In harness engineering, what two components make up an AI agent?

An AI agent consists of the model plus the harness (Agent = Model + Harness).

Why is there a natural trust barrier when working with AI models?

AI models are non-deterministic by nature because they think in tokens, which context and harness engineering help close.

What are the two goals of a well-built outer harness?

It increases the probability the model gets it right first time and provides a feedback loop for it to self-correct.

What is the difference in execution between computational and inferential controls?

Computational controls are deterministic and fast (run by the CPU), while inferential controls are slower, costly, and more non-deterministic (run by the GPU or NPU).

What does "keeping quality left" mean in software development?

It means catching and fixing problems when you write the code because doing it in production is more expensive.

What factors make a codebase more harnessable?

Strong typing, clear module boundaries, and frameworks that hide details from the agent.

What different types of harnesses appear based on usage?

Maintainability harness, architecture fitness harness, and behavioral harness.

What are the three steps of controls?

Measurements, comparison, and corrective action against a standard or goal.

How do feedforward and feedback controls differ in when they act?

Feedforward acts before problems happen to prevent them, whereas feedback acts after results occur to adjust based on outcomes.

Why do systems use feedforward, feedback, and concurrent controls together?

They are all used because they cover each other's weaknesses.

What is an architecture fitness function?

An automated check that tells you if architectural constraints like speed, security, or clean structure are met.

According to evolutionary architecture, how should architecture qualities be handled instead of being written in a document?

They should be tested continuously like code.

In a harness, how does a guide differ from a sensor?

A guide is feedforward: it steers the agent before it acts, with rules, permissions, and context. A sensor is feedback: it watches after the agent acts, with tests and linters, so the agent can self-correct.

What is feedforward's weakness, and what is feedback's weakness?

Feedforward needs good predictions, and you cannot foresee everything. Feedback uses real results, but the damage is already done and there is a delay before you can correct it.