2023-2024 Sem I
This course introduces cloud infrastructure. Students should feel more comfortable with building cloud services after having done this course.
- Prerequisites: COL331 or equivalent.
Note: The course includes programming assignments and thus expects proficiency with systems programming and debugging.
- Credits: 3-0-2
- Slot: AA, Mondays and Thursdays 2-3:15pm in LH521.
- Abhisek Panda: csz202445 AT cse.iitd.ac.in
- Ravi Ranjan Singh: jcs222659 AT csia.iitd.ac.in
- Reading material: There is no textbook for the course. Most lectures will link to more reading material. Lecture notes can be found here and here.
- Acknowledgements: Thanks to Robert T. Morris, MIT and Mythilli Vutukuru, IITB; parts of this course have been inspired by courses made available by them.
- 30% labs (programming assignments)
- 20% project
- 10% quizzes
- 20% minor exam
- 20% major exam
- Labs are to be done on Baadal. You will need VPN access to IITD network!
- Discussions should be done on Piazza.
- Audit criteria: 40% or more marks in total. 40% or more marks in major+minor exams.
- Ethics: Please re-read IITD honour code. Cheating will get an F in the course. Why should I not cheat?
- Late policy: To help you cope with unexpected emergencies, you can hand in your Labs solutions late, but the total amount of lateness summed over all the lab deadlines must not exceed 72 hours. You can divide up your 72 hours among the labs however you like; you don’t have to ask or tell us. You can only use late hours only for Labs.
- There will be no make up quizzes. We will count the scores from your best (n-1) quizzes where n is the total number of quizzes.
- Translate existing programs to distributed system. (Distributed shared memory)
- Batch computation (MapReduce, Spark), streaming computation (Spark streaming, Flink, Google Dataflow), ML training (Tensorflow)
- The problem of late data in streaming computation (Millwheel, Google dataflow): watermarks, triggers, windows.
- Fault tolerance strategies: re-run deterministic idempotent functions (MapReduce, Spark), asynchronous consistent checkpoints (Flink), inconsistent checkpoints (TensorFlow).
- Straggler mitigation, scalability, locality, etc.
- PACELC theorem: If partitioned, choose between availability and consistency, else choose between latency and consistency.
- CP systems:
- Linearizability. Raft: quorums, leader election.
- Serializability. Google Spanner: distributed transactions, TrueTime, hybrid logical clocks.
- AP systems:
- Amazon dynamo: eventual consistency, hashing, gossip protocols, dotted version vectors, conflict-free replicated data types (CRDTs)
- Somewhere between CP and AP
- Google file system
- RedBlue consistency
- CPU virtualization: KVM, Popek-Goldberg theorem
- Memory virtualization: 2-d page tables
Disclaimer: Actual course contents may differ slightly depending on student interest. Reach out to the instructor as soon as possible if there is a particular interest in a topic.
LEC 1: Introduction.
LEC 2: Scalability, Task DAGs, FaaS.
Ch.5 of Introduction to Parallel Computing
LEC 3: Struggles with DSM.
LEC 4: MapReduce. Release Lab 1.
LEC 5: Spark: Resilient distributed datasets.
LEC 6: TensorFlow operational semantics.
Lab 1 due
LEC 7: TensorFlow.
LEC 8: CIEL. Release Lab 2.
LEC 9: CIEL. Quiz 1.
LEC 10: Spark streaming.
Lab 2 due
LEC 11: Virtual time and global states.
LEC 12: Flink. Quiz 2.
Mental health discussion. Release Lab 3.
LEC 13: Dataflow model
Lab 3 due
LEC 14: GFS. Discuss Lab 3
LEC 15: GFS
LEC 16: Raft. Release Lab 4
Wednesday in-lieu of holiday on 28th
LEC 17: Raft. Release project.
Lab 4 due
LEC 18: Zookeeper. Quiz 3.
LEC 19: CRAQ. Dynamo.
LEC 20: Dynamo. Release Lab 5.
LEC 21: Dotted version vectors.
No class day
LEC 22: CRDT.
Lab 5 due
LEC 23: Spanner.
LEC 24: Spanner. Why virtualization? Quiz 4.
LEC 25: Hardware-assissted virtualization