1. Distributed Systems - Comp 655
Processes - outline
Process & thread basics
Multi-threaded clients
Server basics
Server design issues
Code Migration
11/13/2018
Distributed Systems - Comp 655

2. If you remember one thing…
Threads execute
Processes don’t
A thread owns Id
Program counter
Scheduling priority
Stack
Local storage
A process owns Id
Thread table
Page tables
Authenticated identity
IO resources
11/13/2018
Distributed Systems - Comp 655

3. Distributed Systems - Comp 655
Threads in Java
Threads are objects
… like almost everything in Java
To make one,
Define a class that
Extends java.lang.Thread, or
Implements java.lang.Runnable
At runtime, create and initialize an instance
Call the start method:
mythread.start(); or
new Thread(myrunnable).start();
11/13/2018
Distributed Systems - Comp 655

4. Distributed Systems - Comp 655
Java thread example
public class Worker implements Runnable {
private long myInterval;
private int myReps;
public Worker(long interval, int repetitions) {
myInterval = interval;
myReps = repetitions; }
public void run() {
String me = Thread.currentThread().getName();
for( int repsDone=1; repsDone<=myReps; repsDone++) {
System.out.println(me+" ["+(repsDone)+"]");
try{ Thread.sleep(myInterval); }
catch(Exception e) {}
11/13/2018
Distributed Systems - Comp 655

5. Starting the thread example
Public static void main(String names[]) {
for( int i=0; i<names.length; i++ ) {
Worker w = new Worker(1000,7);
Thread t = new Thread(w);
t.setName(names[i]);
t.start(); }
11/13/2018
Distributed Systems - Comp 655

6. Distributed Systems - Comp 655
Possible output
java ThreadExample hey whats up
hey [1]
whats [1]
up [1]
hey [2]
whats [2]
up [2] …
11/13/2018
Distributed Systems - Comp 655

7. Distributed Systems - Comp 655
Why threads?
Sometimes you have to block
usually for IO
Context switching (for processes) is expensive
Why is context switching expensive?
Memory management
Why do memory management?
Concurrency transparency
11/13/2018
Distributed Systems - Comp 655

8. Distributed Systems - Comp 655
Context switching
The most expensive
operations
11/13/2018
Distributed Systems - Comp 655

9. Lightweight processes
User-level thread packages can do context switching for threads MUCH less expensively
BUT, if one thread makes a blocking system call, all threads block
Solutions:
Lightweight processes
Scheduler activations
11/13/2018
Distributed Systems - Comp 655

10. Lightweight processes
thread
table
Most thread switching is done in user space
LWP switching occurs only when a thread makes a blocking system call
Scheduling routine, thread table, and thread table mutexes are shared by LWPs and the user-level thread package
11/13/2018
Distributed Systems - Comp 655

11. User threads multiplexed

12. Distributed Systems - Comp 655
Processes - outline
Process & thread basics
Multi-threaded clients
Server basics
Server design issues
Code Migration
11/13/2018
Distributed Systems - Comp 655

13. Multi-threaded clients
Typical web browser
Main thread gets the page you requested
Additional threads are dispatched to fetch images, frames, scripts, etc
User can start viewing before all the data has been received
Parallelism can shorten elapsed download time when web server is replicated
11/13/2018
Distributed Systems - Comp 655

14. Distributed Systems - Comp 655
Processes - outline
Process & thread basics
Multi-threaded clients
Server basics
Server design issues
Code Migration
11/13/2018
Distributed Systems - Comp 655

15. Multi-threaded server
Java RMI provides a dispatching thread for a Remote object, and a pool of threads to execute calls.
11/13/2018
Distributed Systems - Comp 655

16. Multi-threaded server – plan 1
Thread type
Started by
How many
Responsibility
Dispatcher
Main 1
Listen for requests and queue them
Worker
Main or pool manager 7+
Dequeue a request, process it, send results to client, repeat
Pool manager
Watch size of request queue, and start new Workers if average queue size is too large
11/13/2018
Distributed Systems - Comp 655

17. Threads may need synchronization
Why synchronize? (preview of week 6): concurrent access to shared resources
Concurrent writers can corrupt data
Concurrent reader and writer can confuse the reader
Example: workers de-queue and dispatcher en-queues
All operating systems (and some languages) provide synchronization primitives
When you design a multi-threaded program you have to have a synchronization plan/design/architecture
11/13/2018
Distributed Systems - Comp 655

18. Multi-threaded server – plan 1 synchronization
Object to synch on
Threads involved
How it works
Request queue
Dispatcher, Workers
Dispatcher locks object before queuing a request
Dispatcher notifies one Worker after queuing a request
When a running Worker has no request to process, it locks the queue, checks for requests; if found, dequeues it, unlocks queue, processes request
If no requests are waiting, Worker waits on the queue
11/13/2018
Distributed Systems - Comp 655

19. Multi-threaded server – plan 2
Thread type
Started by
How many
Responsibility
Dispatcher
Main 1
Listen for requests and start a Worker to handle each one
Worker
1 per request
Process a request, send results to client, end
Compared to plan 1, no synchronization, but more overhead for starting and stopping threads
11/13/2018
Distributed Systems - Comp 655

20. Distributed Systems - Comp 655
Processes - outline
Process & thread basics
Multi-threaded clients
Server basics
Server design issues
Code Migration
11/13/2018
Distributed Systems - Comp 655

21. Typical server design issues
How to find it
While minimizing use of well-known endpoints (why?)
How to interrupt it
Stateful vs stateless
Object activation policy
11/13/2018
Distributed Systems - Comp 655

22. Distributed Systems - Comp 655
The super-server
11/13/2018
Distributed Systems - Comp 655

23. The super-server: inetd
11/13/2018
Distributed Systems - Comp 655

24. Interrupting a server that is processing a long-running request
Major options
Abrupt termination
Out-of-band control messages (see below)
In-band control messages with priority
Response chunking with option to quit after each chunk
Client
TCP Priority Data Transfer: "Urgent" 
Out-of-band control messages
Server
Regular (data) messages
11/13/2018
Distributed Systems - Comp 655

25. Stateful vs stateless servers
Stateless server forgets everything after each request has been processed. E.g. HTTP server
Stateful servers can have better performance and be less annoying
BUT
At the cost (usually) of losing
Relocation transparency
Failure transparency
11/13/2018
Distributed Systems - Comp 655

26. Object activation policy options
Object creation
Per request
Fixed-size pool
Dynamic pool
Threading
Single
Fixed-size thread pool
Dynamic thread pool
Thread per object
An object adapter is an implementation of an activation policy.
11/13/2018
Distributed Systems - Comp 655

27. Object adapters in an object server
and two activation policies
For example,
new object on a
new thread for
each request
For example,
single object on a
single thread handles
all requests
11/13/2018
Distributed Systems - Comp 655

28. Distributed Systems - Comp 655
Code Migration
What it is: moving programs across a network on demand
11/13/2018
Distributed Systems - Comp 655

29. Distributed Systems - Comp 655
Why migrate code?
Performance improvement
Flexibility
Simplified administration
11/13/2018
Distributed Systems - Comp 655