What happens when many clients want to contact the server at the same time ? The iterative approach consist in “enqueuing” the requests (this is done automatically by the system) and serving them one by one : Accept the next connection and build the socket Read request Process request (send the response) This will inevitably mean that if there is a client with a small request (for example, a small file request) will have to wait until the large requests are over If more than the allowed number of clients at socket queue request a service during this time they are rejected !!! There may be also some delays in attending clients by waiting some information (client are asked to type in something) There are many Disk I/O operations in a file serving scenario which are normally “slow”and do not require network traffic or CPU So there is a lot of resources standing idle with this schema NOTAS

A sequential (iterative) server attending more than a client A SERVER A CLIENT 4444

During the transfer of the file the server cannot hear at the port 4444 for requests A SERVER A CLIENT 4444

Only after the transmition is over the server can attend another request A SERVER A CLIENT 4444

This may involve some delay (typing, transmitting large files) A SERVER A CLIENT 4444

Sometimes a large waiting time for the client may mean a timeot for the connection A CLIENT A SERVER Timeout A CLIENT 4444 A CLIENT

The Concurrent Server A concurrent server can attend many clients at the same time While transferring a file, it can still keep listening for requests Every time a request comes, a new parallel line of statements execution is started for attending this request. After this the server can “hear” again at the server socket for further requests Different approaches have been developed to implement parallel lines of executions within a program The operative system plays an important role NOTAS

In concurrent servers there is a main program which listens for client requests A SERVER A CLIENT 4444

After a client is contacted, a new line of execution is created which will attend the request in parallel. A SERVER A CLIENT 4444

The main program will return to the listening loop while the first client is being attended, so it can accept another request A SERVER A CLIENT 4444

… and create another parallel line of execution for attending the client A SERVER A CLIENT 4444

… so a request of a third client can be immediately be accepted A SERVER A CLIENT 4444

… thus having the three clients being attended in parallel A SERVER A CLIENT 4444

The Concurrent Server Algorithm Master program of the server 1. Create a server socket 2. Wait for and accept a client’s request 3. Create a parallel line of execution to attend the requirement of the new client 4. Goto 2. Slave (parallel) process : 1. Receive the parameters of the communication established by the master (socket or input/output data streams) 2. Attend client’s request (read filename, transfer data) 3. Return (disappear !) NOTAS

Parallel Lines of execution If there is only 1 CPU, why create parallel processes? It often makes server programming easier. Because there are more CPU involved !!!!! The concept of parallel processes implemented at the S.O level appeared with UNIX and C. In C a new process can be created by executing the fork() statement int i = fork() creates an exact copy of the running process. Both continue executing the same program with the same variables. For the “father” the value of i will be the id of the created process. For the child process this value will be 0 When programming concurrent servers the father will be the main process accepting requests and the child will process the client’s request NOTAS

A simplified example of the use of for for implementing concurrent servers main() { int pid, msock, ssock; msock = passivesock(port, “tcp”, qlen); /* see chapter 10.4 from Internetworking with tcp/ip from Douglas Commer for the implementation */ while(1) { ssock = accept(msock, &fsin, &alen); pid = fork(); if (pid == 0) { process client’s request; return; } NOTAS

Problems with fork() in UNIX The creation of a new process is a time costly procedure In some texts it is suggested to create a set of processes at the beginning which will be activated later by the main program. When a client arrives the parameters of the communication are passed vi pipes to child. The new created process duplicates all the variables !!! It is not easy to manage the processes which ended in an abnormal way. They keep the resources !!!. Sometimes is preferred to use the “select” statement: it basically selects from an array of sockets the first one which has available data to read. NOTAS

In JAVA is better to use Threads A thread is a sequence or a flow of instructions which are executed in parallel to the main sequence of the program. It has a start and an end. A thread can only be created and lives inside an already running process. A process can start as many threads as necessary. Because of this, the main program has a better control of the threads it started. They can be created, started, suspended, or reactivated by the program. Threads are implemented as methods of a certain class. When this method (normally called run) is activated, it starts to run in parallel to the method that called it. As any other method, it can define its own set of variables NOTAS

Using threads to implement concurrent servers Main (master) program: Create ServerSocket ss; While(true) { Socket s = ss.accept(); Create new thread with the socket s as parameter; start executing thread; } Define a thread class with a parallel method (run) which implements the processing of the clients request NOTAS

Implementation of Threads One way (perhaps the most clear one) is to define a new class which is an extension of the Thread class and overwrite its run() method. Thread is an existing class This thread class has a run method (originally with no statements) which will be executed in parallel when activated. In order to execute this method a program must create an object of the extended class and apply the start() method to it. The definition of the new Thread class should look like this: public class MyThread extends Thread { And somewhere should appear: public void run() { //instructions to be executed in parallel } NOTAS

Example public class SimpleThread extends Thread { public SimpleThread(String str) { super(str); } public void run() { for (int i = 0; i < 10; i++) { System.out.println(i + " " + getName()); try { this.sleep((int)(Math.random() * 1000)); } catch (InterruptedException e) {} System.out.println("DONE! " + getName()); The this.sleep(miliseconds) must appear in a try and catch block NOTAS

Use of this new class public class TwoThreadsTest { public static void main (String[] args) { SimpleThread t1,t2; t1 = SimpleThread("Jamaica"); t2 = SimpleThread("Fiji"); t1.start(); t2.start() } The start() triggers the execution of the run method of the thread. There are other methods which can be applied to a thread: suspend(), resume(), stop(). These are however deprecated because their use can easily lead to deadlocks NOTAS

Sometimes you cannot program the server as an extension of a thread For example, if the server has to implement a graphical interface it should be declared as an extension of a Frame, if it is programmed as an Applet it must extend the Applet class The we can use the interface Runnable, which means the Server will also be seen as a Runnable class and has to implement the run method run(). To initiate a parallel execution of the run method a Thread object must be created. In the creation a pointer to the server object (this) should be passed as parameter. With this, the Thread object will have access to the run method. The run method implemented in the server class will be executed by performing start() on the new created Thread object. NOTAS

Example with interface Runnable See and execute the program NoSincron.java Note that the Threads created in this way will have access to all resources of the master. With the other form of creating threads, resources from the master object can be made available by passing a point to itself on creation of the thread NOTAS

Critical regions and semaphores in Java Java provides basically two methods for controlling concurrent access to resources You can declare a whole method as a critical region (see sincron1) You can use the semaphore of any object (see sincron2) Another way to use semaphores in java is with wait() and notify() NOTAS

Implementing a concurrent file server (as an extension of Thread) Define a new Class as an extension of the Thread class which has a socket as Object variable. Program a constructor which receives a socket as parameter. Program the run methods to attend the client connected at the other extreme of the socket. Program a main method which consists of an infinite while. Inside the while program the acceptance of a new client, the creation of a new thread with the socket obtained, and the starting of the thread execution. The client does not change, in fact, it does not know if it is being served concurrently or in parallel See: MultiArchServidorThread.java &MultiArchServidor.java NOTAS

Broadcasting a text to many clients Hello Hello Hello Hello Hello Hello

The server: receiving a new client 4444 Client contacts server at 4444

The server: receiving a new client 4444 A new socket is placed and the output channel is opened They are kept in an array

The server: Broadcasting a message Type a message When a message is entered the server will distribute it to the connected clients BraodcastServerNF BroadcastCliente

It is now easy to extend this to a chat system Hello Hello Hello Hello Hello

Conditions for implementing a chat system Server must be listening to requests of new clients AND to messages which are sent by already connected Client must be listening to messages from the server AND to the keyboard for messages the user wants to transmit. There are many approaches for implementing this in TCP/IP No one is the “absolute correct” solution, all them have their advantages and drawbacks Normally (like everywhere in computers) “faster” solutions will use more memory

Solution 1 The server: The client A thread listens to new clients trying to join the chat party When a new clients connects, a PrintWriter and a BufferedReader for that client are created. The print Writer is kept in a vector A thread is created and receives a pointer to the PrintWriters’ vector and the BufferedReader. It reads the input from the client and writes it to all PrintWriters (clients) The client Graphical interface for reading lines from the keyboard (there is a thread which triggers the execution of the actionPerformed method) and sends it to the server A thread will read input from the server and display it on a text area See: Chat1Server.java, Chat1ClientThread.java & Chat1Client

Solution 1 Schema New Client Thread Client3 Client1 PrintWriter BufferedReader

Solution 2 The server: The client The server has only one thread listening at one port for attending all the requests of the clients. The clients contact the server for registering (reg), sending a message (msg) or logging out (des) The server keeps a PrintWiter vector to keep track about which clients participate in the chat party. Additionally, it keeps a vector with the nicknames of the clients The client Graphical interface for reading lines from the keyboard (there is a thread which triggers the execution of the actionPerformed method) and sends it to the server. It also shows the nicknames of the participants A thread will read input from the server. This can be a message (msg) or a refreshment of the login list See: CICChat.java & CICServer.java

Solution 2 Schema New Client Client1 Client2

Solution 2 Client3 Client1 Client2

Peer-to-peer solution for TCP/IP Every program should be client and server at the same time When a new member wants to join the party, he/she shoud contact anyone of the group and ask for the list of contacts After retrieving the list of contacts (hostnames or addresses) he/she should open an input and output channel with everyone (including the one who provided the list)

Peer-to-peer solution: a user starts a chat party by listening on a port for others wanting to join

A user wanting to join contacts the initiator A user wanting to join contacts the initiator. An InputStream and an OutputStream is opened on each

A third user may contact anyone and recover the list of participants, in this case each one has the address of the other in the list

Then contacts all of them opening Input/Output sreams Then contacts all of them opening Input/Output sreams. All them now have two entries on their participant list

Particularities of this implementation The program is written in the ChatPtP.java file In order to allow testing in one computer, it is necessary to give the port number at which the program will listen for newcomers java ChatPtP localport starts a new Chat party java ChatPtP remotehost remoteport localport joins an existing party (on remotehost at remoteport a program is waiting for new members) The is one thread for listening to newcomers Another listens for input on keyboard and sends it to all connected participants There is one thread listening for input for each of the other participants Not very easy, isn’t it ?

Problems of this implementation What happens if one user leaves the chat party ? What happens if during the second connection phase (retrieving the list and contacting the participants) another newcomer joins the party by asking a third one for the list ? This is a very critical problem which has been more or less addressed in some (sometimes hard to understand) papers.

Hausaufgabe 1: ftp client und server (ArchClientRobust & ArchServerRobust ergänzen) Der Client kann Dateien hoch oder runterladen PUT filename GET filename Die Dateien werden durch einen neuen socket geschickt/empfangen Kann auch die liste der Dateien anfordern DIR

Hausaufgabe 2: Erweitrn des Peer-to-peer TCP/IP Chat programs um: (ChatPtP.java) Das Programm soll eine graphische Benutzerschnittstelle haben wie beim CICChat Diese Schnittstelle muss auch zeigen, die “Nocknames” aller Teilnehmern Ein Teilnehmer muss sich von der Gruppe “verabschieden” bevor er die Gruppe verlässt (das will noch längst nicht alle Probleme lösen aber einige schon)

A concurrent web server This will be implemented only for HTML files and classes (like servlets), but it is easily extensible for attending different types of requirements For each client a new thread is created The process is according to the type of the request Httpd (server) HttpProcessor processRequest() thread browser HttpOutputStream HttpFile HttpInputStream HttpClass Echo HttpException HttpClassProcessor

Implementing state in a web server The server must keep tracking of the different users that contact it The client must submit this information The Echo2 class keeps track of a “shopping cart” for different users The client must submit a request like: user=username&product=productcode&qtty=number

Architecture of a generic file server Directory service Aplication Flat file service Client Module

Components Flat File Service: Implements the operations which work directly on the files. It uses a Unique File Identifier (UFID). A new one is generated for each new file Directory Services: is a FFS client , provides a mapping between the UFID and the textual names of the files. It also porvides the necessary functions for managing directories and obtain UFID.Directories are stored as plain files. Client module: Runs in every clinet computer, integrates and extends the FFS and DS operations in an interface application used by programmers. Contains information for localizing files over the network. Provides efficiency by implementing a caché

A model for an FFS interface read(FileId, i, n) : attempts to read up to n bytes from a file starting from the byte in the position i. write(FileId, i, Datos): writes a data sequence starting from the position i into the specified file create() : creates a new file (empty) and returns its UFID delete(FileId) : deletes the file getAttributes(FileId) : returns a structure containing the file attributes setAttributes(FileId, attr) : sets the file attributes according to what is stored in the structure

Access Controls On a local file system it is necessary check the access rights of the file user only when it is opened and the rights are kept until the file is closed On a distributed system the checking are made at the server side. There are two strategies used in order to keep the server stateless: The checking is done when the filename is converted to the UFID and the result is packed as a “capacity” which is returned to the client. The client uses this capacity for each further access. The checking of the user’s rights is made every time the file is accessed. The second one is the most used (in NFS & AFS) because its simplicity

Model for the interface Lookup(Dir, File) localises the name of the file in the directory UFID AddName(Dir, Name, File) If Name was not in the directory, the pair(Name,File) is added modifying the corresponding file UnName(Dir, Name) the pair (Name, file) is deleted from the directory getNames(Dir) turns the list of names in the directory

The NFS Application Virtual System Virtual System Sist Local Sist Server NFS Sist Local Sist Local Client NFS

Caracteristics of the NFS Communication is implemented over RPC and is open. Resides in the server’s kernel The identification of the files is by file handlers containing following information: Filesystem identifier i-node number or file i-node generation number The “state” is kept in the client in a v-node Client authentication is done in every access.Client provides ID and group ID Flat file & directory services are integrated The mount service provides a link to a remote system

Cache in NFS Unix provides standard Cache mechanisms: buffer cache, read ahead, delayed write NFS Cache on Client’s side: data for writing are stored in the cache memory and are written when a commit takes place (buffer full or closing the file) NFS Cache on server’s side: results form read, write, getattr, lookup and readdir are stored locally. This can introduce some inconsistencies with the versions stored at the different clients’ machines because writings in one client are not distributed at the moment to the others. Clients are responsible for maintaining their caches updated. This is done with the help of timestamps: Tc= time of last synchronization of the cache, Tm= time of modification At a certain time T the cache will be still valid if (T - Tc < t) o (Tmcliente = Tmserver). Normally t will be 3-30 secs for files and 30-60 for directories

The AFS Aims to a better performance in situations of scalability Principles Whole-file serving: the content of the whole file is transferred to the client (even if the client has requested a small part of it) Whole-file caching: The file transferred are stored in the local cache memory. The cache is almost permanent. Procedure When the client opens a remote file, the whole content is ttransferred if it was not there already Read/write operation are done locally With a close, a copy of the file is transmitted to the server

The AFS Architecture Application Unix Kernel Vice Local Sist Venus Unix Kernel

Consistency of the Cache Every time a file is transmitted from the server to a client a callback promise is provided which guarantees that if other client modifies the file, this one will be notified The callback status can be either valid or cancelled When the file is transferred to the client the callback is put on valid. When a callback is received from the server (another client did modify the file) the callback promise is put on cancelled Every time the client wants to open a file, it searches it first in the cache. It if is there the callback promise status is looked, if it is still valid, the cache is used by the client, if it is not there or the callback is cancelled, a new version is transferred from te server If the client’s computer reboots,it asks for a timestamp for every file in the cache to the server. If it is consistent with the local timestamp the callback is put on valid, if not on cancelled

Transmitting Objects via TCP Transmission: marshalling, delivery & unmarshalling. The key for this is the Object Serialization: convert the into a representation which can be transmitted across the net (String) All native Java Objects are serializables. For user defined objects it is necessary to write implements Serializable in their definition (this does not include static variables or references to local things such as files or sockets)

Transmitting Objects via TCP The classes which allows the transmit: ObjectInputStream readObjetct() ObjectOutputStream writeObject() The user can change the “standard” serialization mechanism by declaring implements Externalizable This means, the user must implement Void writeExternal(ObjectOutputStram o) Void readExternal(ObjectInputStream i);