CMSC 818e Day 1

Reading time ~3 minutes

These are notes taken during CMSC 818e: Distributed And Cloud-Based Storage Systems. Course webpage and syllabus here.

Day One

Readings

There are readings for each class; do the readings before class and submit notes via blog by 6PM the night before the reading is due.

Projects

  • Project 1: A Key-Value Store in Go - implement the second half of a remote procedure call (Due midnight Sunday Sept 9th)
  • Project 2: Serialization, Persistence, and Immutability
  • Project 3: Replication, and a Cloud Client
  • Project 4: Distributed Consensus
  • Project 5: Building High-Level Abstractions from a Shared Log

Access projects via CMS Gitlab; submit via Dr. Keleher’s server.

Use Piazza to talk about projects; for Go troubleshooting.

Distributed Systems

Decentralized:

  • no single master (usually), point-of-failure
  • local control/autonomy
  • some have done unusual things like using entropy to decide

(often)P2P

  • all are peers
  • peers are basically the same
  • file-sharing:
    • sometimes means data distribution, control central

Distribution and Geo-replication:

  • distance (latency, no simultaneity, agreement difficult)
  • “a DS is one where networked computers coordinate activity only by passing messages”
    • Internet, intranets, mobile devices
  • small
    • low latency
    • high bandwidth
    • secure, reliable interconnect
    • dependent failure
    • coordinated resources
  • big
    • high latency
    • low bandwidth
    • autonomous
    • unreliable network
    • fear and distrust (byzantine failure)
    • independent failures
    • decentralized administration

Challenges

  • private communication over private networks
    • who sent it (auth), integrity, privacy
  • building reliable systems from unreliable components
    • independent failures

“A distributed system is one in which the failure of a machine I have never heard of can prevent me from doing my work. - Leslie Lamport

  • location
    • placement for efficient sharing
    • finding it later
  • coordination and shared state
    • what should we do, and when?
    • can we agree on what we’ve done?
    • who is this “we” you speak of?

Reliability

  • recoverability:
    • don’t lose data if a failure occurs (also durability)
    • assume
      • nodes have storage
      • volatile: fast (memory)
      • non-volatile: slow (disks)
    • but NVRAM getting cheaper, flash memory, etc
  • availability
  • survivability
  • also
    • security, adaptability, agility, etc

Consequences of distribution

  • concurrent processes
    • user work independently
    • non-determinism, race conditions, sync, deadlock, liveliness
  • no global clock
    • coordination through messages (CSP)
    • clock sync works, but only to a degree
  • no global state
    • no single process has knowledge of entire state of the system
  • failures
    • node failures
    • network partitions

Why Go

  • interfaces - no inheritance, no objects in Go, just structs
  • slices > array
    • slices give you a pointer
    • copy by value => copy by reference
    • append
  • type safe
  • gothreads - lightweight
  • channels - the way that gothreads talk to each other. FIFO buffer, synchronize multiple writers and readers. can be single value, can be buffered
  • garbage collection - pervasive in Go. Can be a problem - will happen asynchronously with respect to your program, don’t know when it will happen

Preliminaries

  • GOPATH
  • go build/install/get/run …
  • emacs? go-mode comes w/ distribution
  • visual studio?
  • atom w/ go pro?

Gotchas

  • everything is passed by values: pointers, but also structs, arrays (use slices)
 var arr = [3]int{1,2,3}

 sl := arr[0:2] // first two items but not last
 fmt.Println(sl)

 sl2 := []int{4,5,6}
 fmt.Println(sl2)

 sl3 := append(sl2, 7, 8)
 fmt.Println(sl3)

 sl4 := append(sl, 7, 8)
 fmt.Println(sl4)
  • no pointer arithmetic
  • there is a ++ operator, but it’s not an expression
  • no type aliases
    • ` type dint int` no distinct types
    • can add methods to user-defined types
  • duck typing - hard to look at a struct and see what type it is
  • no ternary
  • no constructors
  • no destructors
  • no default args
  • everything is a package

The CAP Theorem

The CAP theorem states that it is impossible for a distributed data store to simultaneously provide more than two out of the following three guarantees:

  • Consistency: Every read receives the most recent write or an error
  • Availability: Every request receives a (non-error) response – without guarantee that it contains the most recent write
  • Partition tolerance: The system continues to operate despite an arbitrary number of messages being dropped (or delayed) by the network between nodes

Exercise: Loops and Functions

Practice with control flow: Fill out the definition of the square root iteratively using the Tour of Go implement a for loop:

// Modify the code below to compute a square root w/ the iterative approach
// outlined. You may use "math.Abs()", but not any other math method.
// Modify sqrt to return both a number of iterations, and the final value,
// modify main to print them out.
package main

import (
	"fmt"
	"math"
)

func Sqrt(x float64) (float64, int) {
	var z, diff, iter = 1.0, 1.0, 0
	for i := 1; i < 20; i++ {
		diff = (z*z - x) / (2 * z)
		switch {
		case math.Abs(diff) < 0.01:
			break
		default:
			z -= diff
			iter++
		}

	}
	return z, iter

}

func main() {
	z, iter := Sqrt(50)
	fmt.Printf("After %v iterations, z = %v\n", iter, z)
}

A Parrot Trainer Eats Crow

In this post, we'll consider how it is that models trained on massive datasets using millions of parameters can be both "low bias" and al...… Continue reading

Embedded Binaries for Go

Published on February 06, 2021