wayground logo

Free Printable Worksheets

Font size

S
M
L
XL
Worksheets

Distributed Systems

Total questions: 73

Worksheet time: 39mins

Name
Class
Date
1.

Primary motivation for parallel execution in ML training

a)

To reduce model complexity

b)

To speed up the model training process for large datasets

c)

To minimize GPU memory usage

d)

To simplify hyperparameter tuning

2.

Training ResNet-50 on ImageNet-1K, single GPU

a)

Minutes to hours

b)

Hours to one day

c)

Several days to two weeks

d)

Over a month

3.

Most popular form of in-parallel training

a)

Model Parallelism

b)

Pipeline Parallelism

c)

Data Parallelism

d)

Hybrid Parallelism

4.

Key characteristic ImageNet vs CIFAR-10

a)

Larger dataset size

b)

More classes

c)

Higher image resolution

d)

Greater label complexity

5.

Primary goal of parallel execution

a)

To enable larger models

b)

To reduce data storage costs

c)

To speed up the training process

d)

To improve model accuracy

6.

Fundamental cause of long training times on single node

a)

Limited CPU cores

b)

GPU clock speed

c)

Mismatch between data loading bandwidth and training bandwidth

d)

Small batch sizes

7.

Effect of higher resolution (ImageNet-1K)

a)

Faster convergence

b)

Larger activations → more memory → fewer images per batch

c)

Reduced overfitting

d)

Higher gradient precision

8.

Why single-node training is "too slow"

a)

CPU overheating

b)

Bandwidth mismatch

c)

Insufficient RAM

d)

Software inefficiencies

9.

Direct consequence of high-resolution activations

a)

Increased model accuracy

b)

Fewer images can be trained at once

c)

Faster backward passes

d)

Reduced need for synchronization

10.

Effect of smaller activations (CIFAR-10)

a)

Lower model capacity

b)

Allows more input images at once → higher training bandwidth

c)

Increased regularization

d)

Slower convergence

11.

Why perfect bandwidth matching removes need for parallelism

a)

Eliminates data partitioning

b)

Because distributed training adds overhead, which would dominate

c)

Reduces model complexity

d)

Enables larger batch sizes

12.

How data parallelism addresses mismatch

a)

By compressing data

b)

By reducing model size

c)

By splitting data loading across multiple accelerators

d)

By using mixed precision

13.

Factor besides number of images (ImageNet vs CIFAR)

a)

Label diversity

b)

Higher image resolution

c)

Class imbalance

d)

Augmentation complexity

14.

GPT-3 288 years example illustrates

a)

High hardware costs

b)

Extreme necessity of distributed training for large models

c)

Inefficiency of transformers

d)

Need for model pruning

15.

Why distributed training despite overheads

a)

Speedup outweighs communication overhead

b)

Simplifies debugging

c)

Reduces energy consumption

d)

Improves model generality

16.

In data parallelism, each GPU/worker holds:

a)

A shard of the model and full data

b)

A full copy of the model and a partition of the data

c)

Full data and no model

d)

A partition of both model and data

17.

Purpose of model synchronization:

a)

Reduce memory usage

b)

Ensure all workers' models share the same parameters after each iteration

c)

Compress gradients

d)

Balance workload

18.

Why traditional Gradient Descent (GD) doesn't work in parallelism:

a)

Requires excessive memory

b)

Requires gradients over the full dataset, which no worker has

c)

Too slow

d)

Incompatible with GPUs

19.

How input data is handled across workers:

a)

Replicated identically

b)

Partitioned into disjoint subsets

c)

Randomly sampled

d)

Augmented differently

20.

Term for workers' models diverging:

a)

Gradient explosion

b)

Parameter drift

c)

Model divergence

d)

Synchronization failure

21.

At start of training, each worker:

a)

Has random model subsets

b)

Holds a partitioned model

c)

Holds a full copy of the model, initialized identically

d)

Has no model

22.

Why SGD is necessary in parallelism:

a)

Low memory footprint

b)

Allows updates using mini-batches (local data subsets)

c)

Faster convergence

d)

Better regularization

23.

Two main actions in synchronization:

a)

Aggregate losses & broadcast models

b)

Shuffle data & reduce parameters

c)

Collect gradients from workers & redistribute the aggregated result

d)

Compress activations & synchronize caches

24.

Key difference between training & inference:

a)

Data augmentation

b)

Synchronization is removed during inference

c)

Batch normalization

d)

Loss computation

25.

Purpose of broadcasting aggregated gradients:

a)

Reduce communication costs

b)

Initialize next batch

c)

Ensure all workers update models with identical gradients

d)

Enable asynchronous updates

26.

Why GD cannot be applied:

a)

High variance

b)

Slow convergence

c)

No worker has full dataset for complete gradient computation

d)

Requires double precision

27.

Without synchronization:

a)

Models converge faster

b)

Models diverge → multiple inconsistent models

c)

Gradients vanish

d)

Data partitions overlap

28.

Data parallelism at inference stage:

a)

Increases accuracy

b)

Synchronization phase is removed

c)

Requires larger batches

d)

Slows down prediction

29.

Goal of synchronization (collect + distribute):

a)

Minimize epoch count

b)

Balance data partitions

c)

Keep all workers' parameters identical

d)

Reduce gradient variance

30.

Why parameters diverge after iteration (w/o sync):

a)

Random initialization

b)

Each worker trained on different subsets

c)

Learning rate differences

d)

Hardware variability

31.

Standard practice for initialization:

a)

Different seeds per worker

b)

No initialization

c)

Fixed random seed to align all workers

d)

Layer-wise initialization

32.

Meaning of "on the same page":

a)

Same data batch

b)

Identical hardware

c)

All GPUs maintain identical parameter values post-update

d)

Synchronized clocks

33.

Two primary roles in Parameter Server architecture:

a)

Coordinator & Executor

b)

Trainer & Validator

c)

Parameter Server and Worker

d)

Aggregator & Distributor

34.

First step a worker takes in training iteration:

a)

Compute gradients

b)

Pull latest weights from the server

c)

Load local data

d)

Push gradients

35.

Where final parameter update occurs:

a)

On each worker

b)

On the data shard

c)

On the central parameter server(s)

d)

During synchronization

36.

Component aggregating model updates:

a)

Worker pool

b)

Parameter server

c)

Gradient cache

d)

Synchronization layer

37.

Parameter Server architecture is:

a)

Decentralized

b)

Federated

c)

A centralized model

d)

Peer-to-peer

38.

Primary worker responsibility:

a)

Store global model

b)

Compute gradients on local data partition

c)

Update parameters

d)

Aggregate gradients

39.

Communication pattern server → workers (weights):

a)

Fan-in

b)

Peer-to-peer

c)

Fan-out communication

d)

Broadcast-reduce

40.

Cause of bottleneck at "Push Gradients" stage:

a)

Worker CPU limits

b)

Server must receive gradients from all workers, dividing bandwidth

c)

Network latency

d)

Gradient size

41.

Sharding in Parameter Server context:

a)

Splitting data across workers

b)

Splitting model parameters across multiple servers

c)

Partitioning gradients

d)

Dividing epochs

42.

Communication pattern when all workers send gradients to server:

a)

Scatter

b)

All-gather

c)

Fan-in

d)

Reduce

43.

Negative consequence of more workers (fixed cluster):

a)

Higher per-iteration time

b)

Increased accuracy

c)

Less data processed per iteration

d)

Faster convergence

44.

Negative consequence of more parameter servers (fixed cluster):

a)

Higher memory usage

b)

Reduced model capacity

c)

Fewer nodes compute gradients → lower throughput

d)

Increased latency

45.

Trade-off when choosing #servers vs #workers:

a)

Accuracy vs speed

b)

Communication bandwidth (servers) vs training throughput (workers)

c)

Memory vs storage

d)

Synchronization vs asynchrony

46.

Shortcoming of Parameter Server design:

a)

Low scalability

b)

Poor fault tolerance

c)

High coding complexity (explicit roles & protocols)

d)

Slow convergence

47.

In single server, N workers: why each worker only gets 1/N bandwidth?

a)

Worker GPU limits

b)

Server's bandwidth divided among N workers

c)

Network switches

d)

Gradient aggregation overhead

48.

How sharding solves fan-out bottleneck:

a)

Reduces gradient size

b)

Each worker pulls from N servers concurrently, saturating its link

c)

Compresses parameters

d)

Uses lossy compression

49.

Reason for decline of Parameter Server:

a)

Low accuracy

b)

Difficult to determine optimal server/worker ratio

c)

High cost

d)

Slow inference

50.

How All-Reduce differs from Parameter Server in node roles:

a)

Uses dedicated aggregators

b)

It abandons the parameter server role, making all nodes equivalent workers

c)

Requires more memory

d)

Centralizes gradients

51.

Collective communication primitive for synchronization:

a)

Broadcast

b)

Scatter

c)

All-Reduce

d)

Gather

52.

Purpose of Broadcast collective operation:

a)

Aggregate data

b)

To distribute the same data from one node to all others

c)

Shard parameters

d)

Reduce memory

53.

All-Reduce architecture paradigm:

a)

Centralized

b)

Hierarchical

c)

Decentralized

d)

Client-server

54.

Primary advantage over server/worker ratio problem:

a)

Lower memory

b)

Eliminates the ratio issue since all nodes are workers

c)

Faster data loading

d)

Simpler initialization

55.

Communication pattern of All-Reduce primitive:

a)

One-to-all

b)

All-to-one

c)

All-to-all

d)

Point-to-point

56.

Global batch size in data parallel training:

a)

Per-GPU batch size

b)

Total number of training samples across all GPUs in one iteration

c)

Number of epochs

d)

Gradient accumulation steps

57.

Rule of thumb for learning rate scaling:

a)

Divide learning rate by N

b)

Use constant learning rate

c)

Multiply learning rate by N (number of GPUs)

d)

Square root of N scaling

58.

Term for batch size in single-node training (analogous to global batch):

a)

Micro-batch

b)

Gradient batch

c)

Batch size or mini-batch

d)

Epoch size

59.

Why excessively large global batch size is detrimental:

a)

Increases communication

b)

It can harm model convergence

c)

Causes divergence

d)

Requires more servers

60.

Architectural difference Parameter Server vs All-Reduce:

a)

Data partitioning

b)

Gradient aggregation

c)

Parameter Server = centralized with distinct roles; All-Reduce = decentralized homogeneous roles

d)

Learning rate scaling

61.

Negative outcome of too small global batch size:

a)

Slower convergence

b)

Higher variance

c)

Wastes distributed computational resources

d)

Memory overflow

62.

Why All-Reduce has lower coding complexity:

a)

No gradients needed

b)

Communication is abstracted away by libraries (vs explicit coding in Parameter Server)

c)

Automatic batch sizing

d)

Simplified initialization

63.

Contrast of architectures (Parameter Server vs All-Reduce):

a)

PS uses smaller batches

b)

Parameter Server = centralized, All-Reduce = decentralized

c)

PS faster for small clusters

d)

All-Reduce requires more memory

64.

Which describes location transparency in distributed systems?

a)

Physical location is visible to users

b)

The location of an object is hidden from the user

c)

Objects are immovable

d)

Location defines access rights

65.

Which is NOT a goal of distributed systems design?

a)

Scalability

b)

Fault tolerance

c)

Centralization

d)

Resource sharing

66.

Primary challenge of replication in distributed systems?

a)

Storage cost

b)

Maintaining global synchronization across replicas

c)

Network latency

d)

Data encoding

67.

Why administrative scalability challenging?

a)

Conflicting policies regarding usage, management, and security across domains

b)

Limited IP addresses

c)

Hardware heterogeneity

d)

Software licenses

68.

Example of a shared resource in a distributed system?

a)

Local CPU cache

b)

Private SSD

c)

Cloud-based shared storage

d)

GPU memory

69.

Transparency hiding object movement during use?

a)

Access transparency

b)

Relocation transparency

c)

Migration transparency

d)

Concurrency transparency

70.

"Layered architecture" refers to:

a)

Physical network topology

b)

Organization of components into hierarchical levels for specific functionalities

c)

Data replication tiers

d)

Security clearance levels

71.

Main difference stateless vs stateful servers?

a)

Stateless use UDP

b)

Stateful servers keep track of client status, while stateless do not

c)

Stateful are faster

d)

Stateless scale better

72.

Primary benefit of stateful servers?

a)

Lower bandwidth

b)

High performance through caching and prefetching

c)

Easier debugging

d)

Simpler load balancing

73.

How sharding mitigates bottlenecks.

4 lines