CSS434: Parallel

1
CSS434 Operating System Support Textbook Ch6
Professor Munehiro Fukuda
2
Outline

• Processes
• Threads
• Pthread
• Java Thread
• Multithreaded client and server design
• The role of OS/Network for RPC
• Microkernel

3
System Layers
4
ProcessesDefinition and Aspects

• An environment to execute a program
• A process consists of
• A CPU register set (including program counter)
• An dependent address space including
• Text (code)
• Data (global data)
• Stack (local variables)
• Heap (dynamic data)
• Files
• Communication resources (sockets)
• Threads and their synchronization facility

socket
5
ProcessesCreation and Resource Sharing

• A parent process creates child processes through
  fork( ).
• A child process overloads a new program on it
  through execve( ).
• Execution
• They run in concurrent.
• A parent may wait for the termination of a child
  through wait( ).
• Resource sharing
• Resource inherited by children file descriptors,
  shared memory and system queues
• Resource not inherited by children address space

shared
shared
socket
shared
6
ProcessesCreation and Copy-on-Write

• Upon creating a child process
• A virtual address space is allocated to a child.
• Corresponding physical memory is still mapped to
  its parent.
• Upon a write operation, physical memory is
  allocated to a child and data is copied.

Physical address space
Parents virtual address space
data
Shared memory
w
data
stack
text
Shared memory
w
Childs virtual address space
data
w
text
Shared memory
w
stack
stack
text
7
ProcessesCreation and Load Balancing

• Transfer policy Create a new process
• Locally
• Remotely ? Location Policy
• Static transfer processes to predefined
  destinations
• Adaptivetransfer processes to destinations based
  on run-time information
• Centralized load-sharing A load manager take
  care of correcting info and migrating processes.
• Hierarchical load-sharing A system consists of a
  tree structure where each node takes care of its
  child processes for load balancing
• Decentralized load-sharing
• Sender-initiated A heavy-loaded node sends out a
  new process.
• Receiver-initiated A light-loaded node
  advertises its existence.

8
Threads

• Threads
• The basic unit of CPU utilization
• Control of Program
• Process can have two or more parallel controls of
  program ? multiple threads.
• Belong to the same process.
• No protection between threads.
• Advantages
• Light creation
• Light context switch
• Suitable to parallel computing
• Natural form of resource sharing

9
Threads v.s. Multiple Processes

• Maybe, we can implement a server of multiple
  processes?

10
Thread ImplementationLibrary and Class
11
Thread ImplementationC Example
include ltiostreamgt include ltstringgt using
namespace std include ltpthread.hgt include
ltunistd.hgt void thread_func( void param )
for ( int i 0 i lt 5 i ) sleep( 2 )
cout ltlt "I'm a slave " ltlt ( (string )param )
ltlt endl return NULL void main( int
argc, char argv ) pthread_t child
string arg cout ltlt "enter message " cin
gtgt arg pthread_create( child, NULL,
thread_func, (void )arg ) for ( int i 0 i
lt 10 i ) sleep( 1 ) cout ltlt "I'm a
master " ltlt arg ltlt endl pthread_join(
child, NULL ) cout ltlt "Master synched with
slave" ltlt endl
Compilation and Execution
Source Code
g thread.cpp -lpthread a.out enter message
hello! I'm a master hello! I'm a slave
hello! I'm a master hello! I'm a master
hello! I'm a slave hello! I'm a master
hello! I'm a master hello! I'm a slave
hello! I'm a master hello! I'm a master
hello! I'm a slave hello! I'm a master
hello! I'm a master hello! I'm a slave
hello! I'm a master hello! Master synched with
slave
12
Thread ImplementationJava Example
MyThread.java
public class MyThread public static void
main( String args ) String arg
args0 ThreadFunc child new ThreadFunc( arg
) child.start( ) for ( int i 0 i lt 10 i
) try Thread.sleep( 1000 )
catch ( InterruptedException e )
System.out.println( "I'm a master " arg
) try child.join( ) catch (
InterruptedException e ) System.out.println(
"Master synched with slave" ) public
class ThreadFunc extends Thread public
ThreadFunc( String param ) this.param
param public void run( ) for ( int
i 0 i lt 5 i ) try Thread.sleep(
2000 ) catch ( InterruptedException e )
System.out.println( "I'm a slave "
param ) String param
Compilation and Execution
ls MyThread.java ThreadFunc.java javac
MyThread.java java MyThread hello! I'm a
master hello! I'm a slave hello! I'm a master
hello! I'm a master hello! I'm a slave
hello! I'm a master hello! I'm a master
hello! I'm a slave hello! I'm a master
hello! I'm a master hello! I'm a slave
hello! I'm a master hello! I'm a master
hello! I'm a slave hello! I'm a master
hello! Master synched with slave
ThreadFunc.java
13
Client and Server with Threads

• Client
• Can avoid a block on RPC by having two threads.
• Server
• Assume A request consists of 2ms processing and
  4ms disk I/O.
• Single-threaded Server
• Needs 6ms for a request
• Can process 166 request/second
• If three threads pick up and work on a request in
  turn
• Can overlap processing and I/O
• Can process 1000/4 250 request/second.

14
Multithreaded ClientsArchitecture

• Why Multithreaded Clients
• Hide network latency
• Main Thread
• Interacts with a client user
• Work on computation
• Dispatch a request to a child thread
• Child threads
• Sends a request to a given server through TCP or
  RPC.
• Waits for a response
• Forwards a response to the main.

Client
Main thread
Child Threads
User
Pick up
response
RPC or TCP
Server 4
request
Enqueue
15
Multithreaded ClientsExample
Web browser

• Web browser
• Requests and reads the main HTML file.
• For each ltimg srcgt, ltobjectgt, ltappletgt, etc.,
• Spawns a child thread
• Child threads
• Sets up an HTTP connection
• Reads each web part.

Main thread
Child Threads
HTTP
User
Request the main HTML
Server
lthtmlgt
Scan the file
HTTP
Server
Spawn a thread
img
HTTP
Spawn a thread
Server
applet
16
Multithreaded ServersArchitecture

• Dispatcher-worker model
• The worker pool (a)
• Per-request threads (b)
• Per-connection threads (c)
• Team model
• Per-object threads (d)
• Pipeline model (e)

(a)
(b)
Pick up a request
Spawn for each request
accept
accept
queue
(c)
(d)
(e)
Session
accept
request
accept
Spawn for each session
request
process
request
response
17
Multithreaded ServersExample

• Object Server
• Instantiates remote objects
• Associates each with an independent thread
• Let a thread maintain and protect its object
• Requests
• Accepted at request dispatcher
• Forwarded to an object wrapper including objects
  residing under the same policy
• Picked up by the destination thread

Thread Group
Daemon thread may Server for per-object threads
18
Thread/OS Interaction in RPC
19
OS and Network Interaction in RPC
Why does this gap occurs if a RPC arguments grow
beyond a packet size?
20
Monolithic kernel and microkernel
Mach, Chorus modularity, portability,
and extensibility
Unix, Sprite non-modular way,
intractable WindowsNT layering and OO design
but still massive.
21
Role of the microkernel
Microkernel facilitates only address spaces,
threads and local inter-process
communication. Middleware can use directly
Microkernel for better performance or system
processes for convenience and flexible operations
22
Exercises (No turn-in)

• Textbook p262, Q6.7 Explain the advantage of
  copy-on-write region copying for Unix, where a
  call to fork is typically followed by a call to
  exec. What should happen if a region that has
  been copied using copy-on-write is itself copied?
• Why can threads perform their context switch
  faster than processes?
• What are the thread groups? What are the daemon
  threads? How can those contribute to the
  multithreaded server?
• If you comment out the following statement from
  the C code on Slide 11, what change will you
  observer in the execution?
• pthread_join( child, NULL )
• If you comment out the following statement from
  the Java code on Slide 12, what change will you
  observe in the execution?
• child.join( )
• Textbook p262, Q6.8 A file server uses caching,
  and achieves a hit rate of 80. File operations
  in the server costs 5ms of CPU time when the
  server finds the requested block in the cache,
  and take an additional 15ms of disk I/O time
  otherwise. Explaining any assumptions you made,
  estimate the servers throughput capacity
  (average requests/sec) if it is
• Signle-threaded
• Two-threaded, running on a single processor
• Two-threaded, running on a two-processor
  computer.
• Textbook p263 Q6.16 Network transmission time
  accounts for 20 of a null RPC and 80 of an RPC
  that transmits 1024 user bytes (less than the
  size of a network packet). By what percentage
  will the times for these two operations improve
  if the network is upgraded from 10Mbps to 100Mbps.

23
Exercises (No turn-in)

• Q8. The following server code assumes that each
  client program asks three user inputs
• (1) the id of a file to operate on 0 or 1,
  each corresponding file0 and file1
• (2) a file operation type r as a file read
  or w as a file write
• (3) if the file operation is w, a 100-byte
  message to be read from a keyboard.
• The client sends those inputs to the server
  through a socket. If the file operation type is
  r, it receives a 100-byte content of the
  corresponding file from the server and prints it
  out.
• The server spawns two child threads
  child_thread0 and childe_thread1, each
  associated with socket descriptor sd0 and sd1
  respectively. Every time the server accepts a new
  socket request from a client, it receives a file
  id from the client, and passes this socket
  request to the corresponding thread. The thread
  opens its file, checks a file operation type, and
  reads/writes the file according to the type.
• Q8-1. Which server model was used in this tcp.cpp
  program, worker pool, per-request threads,
  per-connection threads, per-object threads, or
  pipeline model? Choose one of these models and
  justify your choice.
• Q8-2. The server in the above program is not so
  scalable. In other words, it cant handle many
  client requests concurrently. Why? Explain the
  reason.
• Q8-3. Which server model(s) are scalable? If you
  change this program into such server model(s),
  what side effect will occur? Describe the major
  problem you have conceived.

24
Exercises (No turn-in)

• include "Socket.h"
• include "pthread.h"
• define PORT 10000
• define SIZE 100 // message
  size
• int sd2 // socket
  descriptor
• void thread_func( void arg ) // thread to
  read and write a given file
• int id (int )arg // my thread
  id
• char fileName (id 0) ? "file0" "file1"
  // if I'm 0, operate on file0
• while ( true )
• if ( sdid NULL_FD ) // check if a
  new socket request came to me.
• continue
• int fd open( fileName, O_RDWR )// open my
  file
• char op // a file
  operation type 'r' or 'w'
• read( sdid, op, 1 ) // read an
  operation type from the socket
• if ( opid 'r' ) // a read
  operation
• read( fd, message, SIZE ) //
  read 100 bytes from my file
• write( sdmyId, message, SIZE ) //
  send them back to the client
• else if ( opid 'w' ) // a write
  operation
• read( sdmyId, message, SIZE ) //
  receive 100 bytes from the client

25
Exercises (No turn-in)

• int main( int argc, char argv )
• int sd NULL_FD // socket
  descriptor
• char id 0 // a file
  (thread) id
• if ( argc 1) // I'm a
  server
• pthread_t child2
• for ( int i 0 i lt 2 i ) // create
  two child threads
• sdi NULL_FD
• pthread_create( child0, NULL,
  thread_func, (void )i )
• while ( true ) // keep
  receiving a client socket request
• if ( ( sd sock.getServerSocket( ) )
  NULL_FD )
• return -1
• read( sd, id, 1 ) // receive
  a file (thread) id
• while ( sdid ! NULL_FD )
• // wait
  for cihld_threadid to be ready
• sdid sd // pass
  this client socket to the thread