Schedule

ADS lectures are held on Friday 6-8 (12:55-15:40) at 312, East Upper Hall (东上院312).
Date Lecture Pre-course reading Paper review
2.23 Lec.1
Introduction & Distributed Systems
Lecture Slides
3.01 Lec.2
Sequential Consistency
Lecture Slides
Memory Coherence in Shared Virtual Memory System
3.08 Lec.3
Eventual Consistency
Lecture Slides
Don't Settle for Eventual: Scalable Causal Consistency for Wide-Area Storage with COPS 1. Making Geo-Replicated Systems Fast as Possible, Consistent when Necessary
3.15 Lec.4
Recovery & Logging
Lecture Slides
Reimplementing the Cedar File System Using Logging and Group Commit Q1 2. The FuzzyLog: A Partially Ordered Shared Log
3.22 Lec.5
Concurrency Control
Lecture Slides
A Critique of ANSI SQL Isolation Levels Q2 3. Transactional storage for geo-replicated systems (PSI)
3.29 Lec.6
Distributed Commit
Lecture Slides
Sinfonia: A New Paradigm for Building Scalable Distributed Systems Q3 4. Building Consistent Transactions with Inconsistent Replication (TAPIR)
4.05 Lec.7
Distributed Consensus
Lecture Slides
Paxos Made Simple Q4 5. In Search of an Understandable Consensus Algorithm (RAFT)
4.12 Lec.8
Distributed File Systems
Lecture Slides
The Google File System Q5 6. TFS: A Transparent File System for Contributory Storage
4.19 Lec.9
Data-parallel Programming
Lecture Slides
MapReduce: Simplified Data Processing on Large Clusters Q6 7. Ray: A Distributed Framework for Emerging AI Applications
4.26 Lec.10
Invitation Talk
5.03 Lec.11
Graph-parallel Computation
Lecture Slides
Distributed GraphLab: a Framework for Machine Learning and Data Mining in the Cloud Q7 8. PowerGraph: Distributed Graph-Parallel Computation on Natural Graphs
15. Gemini: A Computation-Centric Distributed Graph Processing System
5.10 Lec.12
Distributed Data Partitioning
Lecture Slides
PowerLyra: Differentiated Graph Computation and Partitioning on Skewed Graphs Q8 9. Exploring the Hidden Dimension in Graph Processing (Cube)
10. Pragh: Locality-preserving Graph Traversal with Split Live Migration
5.17 Lec.13
Fault-tolerance for Computation
Lecture Slides
Replication-based Fault-tolerance for Large-scale Graph Processing Q9 11. Zorro: Zero-Cost Reactive Failure Recovery in Distributed Graph Processing
5.24 Lec.14
Graph Neural Network Training
Lecture Slides
FlexGraph: A Flexible and Efficient Distributed Framework for GNN Training Q10 12. P3: Distributed Deep Graph Learning at Scale
13. GNNLab: A Factored System for Sample-based GNN Training over GPUs
5.31 Lec.15
Serverless/FaaS
Lecture Slides
No Provisioned Concurrency: Fast RDMA-codesigned Remote Fork for Serverless Computing 14. Serialization/Deserialization-free State Transfer in Serverless Workflows with RDMA-based Remote Memory Map
6.07 Lec.16
Review (12:55 - 14:00)
Lecture Slides

Paper & Questions

Lec.3 Question (Do not need to submit.)

Paper: Don't Settle for Eventual: Scalable Causal Consistency for Wide-Area Storage with COPS
Suppose an application client at data center D1 writes object x with version 2 (x_2) and then object y with version 3 (y_3). Suppose y_3 has propagated from data center D1 to data center D2 but x_2 has not yet arrived at D2. Suppose another application client data center D2 has just read Y_3, is it possible that it might read x_1 next? (If not, why not?) Will the client be blocked waiting for x_2 to arrive from D1? (If not, why not?)

Lec.4 Question.1

Paper: Reimplementing the Cedar File System Using Logging and Group Commit
At the end of Section 4, the paper says that during a one-byte file create FSD writes the leader+data page synchronously to the disk, but records the update to the file name table in memory and only writes it back to disk later. Why do you suppose the FSD designers decided to write the data page synchronously? What (if anything) might go wrong if FSD instead wrote the file's data in the in-memory disk cache, and only wrote it to disk later?

Lec.5 Question.2

Paper: A Critique of ANSI SQL Isolation Levels
Snapshot isolation (SI) differs from serilizatiability due to one anomaly that is possible under SI but not under serilizatiability. Describe the anomality and also give a concrete application for which the anomaly is undesirable.

Lec.6 Question.3

Paper: Sinfonia: A New Paradigm for Building Scalable Distributed Systems
What's the difference between coordinator in mini-transaction's 2PC protocol and standard 2PC protocol?

Lec.7 Question.4

Paper: Paxos made simple
Suppose that the acceptors are A, B, and C. A and B are also proposers. How does Paxos ensure that the following sequence of events can't happen? What actually happens, and which value is ultimately chosen?
A sends prepare requests with proposal number 1, and gets responses from A, B, and C.
A sends accept(1, "foo") to A and C and gets responses from both. Because a majority accepted, A thinks that "foo" has been chosen. However, A crashes before sending an accept to B.
B sends prepare messages with proposal number 2, and gets responses from B and C.
B sends accept(2, "bar") messages to B and C and gets responses from both, so B thinks that "bar" has been chosen.

Lec.8 Question.5

Paper: Google File System
Describe a sequence of events that result in a client reading stale data from the Google File System.

Lec.9 Question.6

Paper: MapReduce
In MapReduce each Mapper saves intermediate key/value pairs in R partitions on its local disk. Contrast the pros and cons of this approach to the alternative of having Mappers directly send intermediate results to R reducers that shuffle and save intermediate results on reducers' local disk before feeding them to the user-defined reduce function.

Lec.11 Question.7

Paper: Distributed GraphLab
How does distributed GraphLab provide consistency in parallel computing, and which consistency is supported by distributed GraphLab?

Lec.12 Question.8

Paper: PowerLyra
Please explain the claim in the paper "For high-degree vertices, the upper bound of increased mirrors due to assigning a new high-degree vertex along with in-edges is equal to the number of partitions (i.e. machines) rather than the degree of vertex".

Lec.13 Question.9

Paper: Imitator
Please name and describe the two alternative recovery mechanisms adopted by Imitator, and discuss their pros and cons.

Lec.14 Question.10

Paper: FlexGraph
Please briefly describe how hierarchical dependency graphs are built in FlexGraph, and point out the specific stage in the NAU abstraction where this process takes place.


Credits: questions and papers from MIT 6.824 and part of slides come from Paul Krzyzanowski (Rutgers), Haibo Chen (SJTU) and et al.