CS 6CC3/4CC3. Advanced Operating Systems. Laboratory 6. O.S. Primitives for BOEWULF and NFS Clustering

Transcription

1 CS 6CC3/4CC3 Advanced Operating Systems Laboratory 6 O.S. Primitives for BOEWULF and NFS Clustering CIS trunk Ethernet mills birkhoff moore FDDI / 1 Gbps ritchie DNS server CAS branch Ethernet CAS branch router CIS DNS servers router/firewall RTOO.syslab ITB232-RTR lion RJ Mbps CISCO 100Mbps switch B1 B2 B3 B4 B5 B6 syslab twig Ethernet (ITB-134) File:lab6cnf.cfl Date:oct04/wfsp LAB to be done in round-robin form. File:4cc04lb6.doc Date:25oct04/wfsp Revision level: 0

2 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-2 INTRODUCTION This lab will introduce you to TCP/IP networking under the UNIX operating system. You will explore various networking facilities available under UNIX using a variety of TCP/IP based applications. You will also implement a client/server application to get a feel for distributed network communication and processing. OVERVIEW This lab is organized into two parts which each require one lab session to complete data collection. Part I will familiarize you with high level network utilities: how they are used and the protocols they employ. FOR THIS LAB GROUP ONLY, since the first session was far to disorganized, this lab will guide you though similar operations but data taking will be much more systematic and time plots of RMS (Record Management System or filesystem) accesses via rcp and cp commands will be made. Hence only Part I need be completed and handed into the marker as one lab report seven days after the end of the second session of the lab (this session). Also the lab will be done from either ITB-236 or ITB-235 under the Solaris machines using X-window access. BACKGROUND Computer networks have revolutionized our use of computers. They pervade our everyday life, from automated teller machines, to airline reservation system, to electronic mail services, to electronic bulletin boards. There are many reasons for the explosive growth in computer networks. The proliferations of personal computers and workstations during the 1990s helped fuel the interest and need for networks. Technology has greatly reduced the cost of establishing computer networks, and networks are now found in organizations of every size. Many computer manufacturers now package networking software as part of the basic operating system. Networking is no longer regarded as an add-on that only a few customers want. It is now considered as essential as a text editor. We are in an information age and computer networks are becoming an integral part in the dissemination of information. Computer systems were initially constructed to be stand-alone entities. Each computer was self-contained and possessed all the peripherals and software required to do a particular job. If a particular feature was needed, such as printer output, a printer was attached to the system. If large amounts of disk storage were needed, disks were added to the system. What helped change this is the realization that computers and their users need to share information and resources. Information sharing can be electronic mail or file transfers. Resource sharing can involve accessing peripherals on another system. Twenty years ago this type of sharing took place

3 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-3 by exchanging magnetic tapes, decks of punched cards, and printer output. Today computers can be connected together from across oceans by electronic networks. A network can be as simple as two personal computers connected via a serial line, or as complex as the TCP/IP Internet, which ties together over millions of systems across the world. The number of ways to connect a computer to a network, are many, as are the various things we can do once connected to a network. Some typical network applications are: Exchange electronic mail with users on other systems. It is almost as common place for address to be found on business cards as are phone numbers. Exchange information and files between systems. Information and data can reside in one place while others access via the network. This avoids having to distribute information via floppies, tapes, or paper. (This mode is often referred to as "sneaker net"!) Share peripheral devices. Efficient use can be made of peripheral devices like printers and tape backup units if they can be shared as a common resource. The cost of the network can be offset through the cost savings returned by the efficient sharing of resources. Distributed processing of computer programs. Networks can be used to distribute a complex application across multiple machines allowing for faster execution. Tasks, which once require the processing power of a Cray, can be handled by a network of smaller cheaper workstations, working in parallel. Remote login and network based file systems. Users can login to interconnected systems without having to physically change terminals. Large numbers of users can often be handle more easily by clustering multiple workstations together and using the network to transparently share a common file space. We will explore the underpinnings for each of these areas further in the lab. LAYERING Given a particular task that we want to solve, such as providing a way to exchange files between two computers, we divide the task into pieces and solve each piece. In the end we connect the pieces back together to form the final solution. We could write a monolithic system to solve the problem, but experience has shown that solving the problem in pieces leads to a better, and more extensible, solution. OSI Model The starting point for describing the layers in a network is the International Standards Organization (ISO) open systems interconnection model (OSI) for computer communications. This is a seven layer model show in figure 1(a). The OSI model provides a detailed standard for describing a network. Most computer networks today are described using the OSI model. Figure 1(b) shows how TCP/IP fits into the OSI model. The transport layer is important because it is the lowest of the seven layers that provides reliable data transfer between two systems. The layers above the transport layer can

4 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-4 assume they have an error-free communications channel. Details, such as sequence control, error detection, retransmission and flow control, are handled by the transport layer and the layers below it. Processes A fundamental entity in a computer network is a process. A process is a program that is being executed by the computer's operating system. When we say that two computers are communicating with each other, we mean that two processes, one running on each computer, are in touch with each other. (a) Layer 7 -- APPLICATION Layer 6 -- PRESENTATION (b) process process OSI layers 5-7 APPLICATION Layer 5 -- SESSION Layer 4 -- TRANSPORT Layer 3 -- NETWORK process TCP UDP ICMP IP IGMP OSI layer 4 TRANSPORT OSI layer 3 NETWORK Layer 2 -- DATA LINK Layer 1 -- PHYSICAL hardware ARP RARP OSI layers 1-2 DATA LINK Figure 1: (a) The Seven Layers of the OSI "Standard"; (b) The TCP Suite of Protocols Relative to this standard. Client-Server Model The standard model for a network application is the client-server model. A server is a process that is waiting to be contacted by a client process so that the server can do something for the client. A typical scenario is as follows: The server process is started on some computer system. It initializes itself, then goes to sleep waiting for a client process to contact it requesting some service. A client process is started, either on the same system or on another system that is connected to the server's system with a network. Client processes are often initiated by an interactive user entering a command to a time-sharing system. The client process sends a request across the network to the server requesting service of some form. Some examples of the type of service that a server can provide are: - Return time of day to the client - Print a file on the server's system - Exchange electronic mail - Read or write to the server's systems filespace - Execute a program

5 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-5 When the server process has finished providing its service to the client, the server goes back to sleep, waiting for the next client request to arrive. We can further divide the server process into two types. 1. When a client's request can be handled by the server in a known, short amount of time, the server process handles the request by itself. We call these iterative servers. A time-ofday service is typically handled as an iterative server. 2. Then the amount of time to service a request depends on the request itself (so that the server doesn't know ahead of time how much effort it takes to handle each request), the server typically handles it in a concurrent fashion. These are called concurrent servers. A concurrent server invokes another process to handle each client request, so that the original server process can go back to sleep, waiting for the next client request. An example of a concurrent server is the sendmail daemon, which accepts from remote sites. Lab Equipment This laboratory can be done using any TCP/IP based network with machines running a variant of the UNIX operation system called Linux. A formal lab has been established with multiple Intel based 586 PCs running either the older Slackware product or the newer Red Hat version of Linux. These machines are interconnected using SMC 10Mbps ethernet cards over a coax based ethernet (thin-net). The PCs have been configured to run X windows and all files, depending on login ID may or may NOT be shared across the machines via NFS. PARTs I & II Introduction The purpose of this lab is to introduce you to the UNIX and TCP/IP networking environment. We assume you have a working knowledge of basic UNIX commands like ls, cd, mkdir, vi, and other basic utilities. In this section you will use network utilities to get a feel for higher level networking functionality. You will use these higher level network utilities to move files through the network from one machine to another and you should make notes and comment on functionality, efficiency, and constraints for each method. Because the data taking is rather extensive, BOTH lab sessions will be used for the following operations. Please try to get as far as possible working your way through the procedures outlined below during Part I (the first session) of the laboratory. Session II will be used to complete the data acquisition. Network Utilities Most flavours of BSD UNIX include a core set of high level TCP/IP utilities. These are refered to as the "r" commands. They include rcp, rsh, rlogin, rdump, rusers, and rup. They, respectively, handle remote copying of files, remote (shell) execution of programs, remote login, remote backups, display of users on remote machines, and display

6 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-6 of uptime statistics of remote machines. Other commands, which are often part of the TCP/IP suite of utilities, are finger, telnet, ftp, and ping. These are used to query user information, log onto remote systems, do file transfers, and check network state of remote systems. The usage of all these commands are explained in online man pages (example "man ping"). Because these instructions require the involved computers to be TRUSTED components (i.e. not asking for login/password for every access), these activities are usually accomplished among computers which are behind a protective firewall. The details of how firewall works and configured are examined in other labs you will be doing later in the course. CIS trunk Ethernet ritchie DNS server FDDI / 1 Gbps birkhoff WOLFs iii ooo FOXes jjj ooo CAS branch Ethernet CAS branch router CIS DNS servers router/firewall RTOO.syslab ITB232-RTR lion RJ Mbps CISCO 100Mbps switch B1 B2 B3 B4 B5 B6 syslab twig Ethernet (ITB-134) File:lab6cnf.cfl Date:oct04/wfsp Figure 2: Architectural topology for the Beowulf clustering laboratory (#6) in cs4cc3/6cc3. Network File System The Network File System (or NFS) was developed by Sun Microsystems to allow files on separate machines to appear to be local to each machine. NFS was designed to allow transparent access to files without regard to where the actual file existed on the network. The NFS system is based on the client/server model and uses TCP/IP protocols for all file accesses. You will notice that regardless of what machine you log into within the Computer

7 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-7 Science Department you will see a consistent home directory. All your files are available from wolfs<n> and foxes<m> and the UNIX PCs in this TCP/IP lab. All UNIX workstations and UNIX PCs use NFS to "mount" users files from birkhoff. In this configuration, birkhoff is referred to as the File Server. See figure 3 for the distinction between file servers and display servers and clients to both, including NFS/local file attachment considerations. Note that for the X-window to function properly, the X-client MUST have authorization to use your console machine as an X-display server so that the.rhosts file must include the wolf and fox computer names for permission to do so. If any errors occur, such as "incorrect login" when the "remote" commands are attempted, contact a technical staff member or TA. Such problems may have occurred in the X-windows lab done individually as the first lab in this course. This may indicate that the machines are not sharing in a "trusted" domain relationship, which can occur when server reboots have recently been necessary to perform. In the case for our BOEWULF machines with have utilized a more global (and therefore highly risky situation but we are behind a firewall so things are OK), method using a file set by root privileges, called hosts.equiv. It can be found in the /etc/ directory of any of the BEOWULF machines and contains the text as shown below. Explain the meaning of the + signs and report on this in the lab write-up. File Transfers There are several ways to transfer files from one machine to another under most TCP/IP networks. (Again an inter-computer trusted environment is assumed) 1. rcp rcp file remote:file rcp remote1:file remote2:file

8 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-8 NOTE: rcp accesses files relative to your home directory on each machine. You must have an account with the same login name on both machines. 2. ftp (available but we will not be using this method here) ftp remote username: password: ftp send file-1 ftp get file-2 ftp quit 3. NFS NOTE: Only superuser can mount filesystems from other machines. All mounts have already been made for you as illustrated in figure 3. Hence a simple copy (cp) command could be employing user transparent NFS (remote) file systems (by previous linked mount), which could be the same as using the explicit rcp command. In this configuration the file path for root, "/" means the local file system and the home file path, "~" means the NFS mount which may or may not be a connection to the local Record Management System (RMS). This will be explored further in the lab session that follows. mount remote:/filesystem /remote cp file-1 /remote/file-2 cp /remote/file-1 file-2 Timing Commands You might wish to time how long it has taken a particular command to execute. You will need to do this to evaluate the efficiency of the various network file transfers you must perform. To time a command you simply pre-append the "time" command to the form of the command. Example: time rcp file remote:file The time taken will display as: 0.050u 0.080s 0: % 0+48k 1+0io 3pf+0w The first number is the amount of USER time taken in ms, the second number is the amount of SYSTEM time taken and the third number will be the amount of elapsed wall clock time. Take particular note of these three times as we shall be recording them and doing plots from a spreadsheet with them. You can ignore the other numbers.

9 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-9 Laboratory Procedures Section 1.0 In the group subdirectory (usually cs4cc3.dat) you'll find ten files named 512b, 1k, 2k, 4k, 8k, 16k, 64k, 512k, 1m, 2m, 4m and 8m. These filenames match the sizes of the respective files. You will be using two of the three file transfer methods mentioned previously to copy each of these files, say, from your home space on rtoo to your local NFS link NFS link NFS link "rtoo" (P266) CPU/RAM cs4cc3 login client cs4cd3 login client RMS cs4cd3 local space cs4cc3 home space "b3" (P75) CPU/RAM cs4cc3 login client cs4cd3 login client RMS cs4cd3 local space "b4" (P75) CPU/RAM cs4cc3 login client cs4cd3 login client RMS cs4cd3 local space lb6fg3.cfl wfsp/oct04 rlogin rtoo -l cs4cd3 rlogin rtoo -l cs4cc3 NFS link X-window display service Legend "WOLF"n - Solaris or "FOX"n - Solaris xtrm xtrm "birkhoff" (SUN-4500) CPU/RAM muss(cas) login client RMS CAS home space AS NFS File Server AS local File Server AS X-Server AS X-client SUN Ultra SPArc 5(10) console login CAS account Figure 3: Login and environment used for lab6 (Beowulf cluster) data extraction. space on the Beowulf (bn) machines. You are to time how long it takes each file to be transferred by each method and you are to plot the time taken in seconds versus the file size for each method. You should copy each file several times noting the transfer times for the first copy and the transfer times for subsequent copies. When plotting show time discrepancies as error bars on the graph. Also provide the OS system and user times as illustrations of the OS load during these transfers. The easiest way to do these transfers is probably to use the setup illustrated in figure 3. Here, a login with one group member s CAS or muss loginid and password will start an X-session. Start up two x-terminals and attach each to the

10 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-10 rtoo machine as two different loginids: cs4cc3/cs4cc3 and cs4cd3/cs4cd3. Next, within each xterm, use the Secure Shell (ssh) method (although a telnet session would work) to login to say the b3 machine. The difference, as noted in the diagram, is that cs4cc3 login has script files on the boewulf (bn) machines that establish a NFS link to rtoo, while the cs4cd3 logins do not. Figure 4: File configuration for the cs4cd3 login on b3 (left) and b4 (right) machines via ssh. The data files to be used for transfers should be found (that is, determine which directory they are in) on b3 for each login ID and also a temporary subdirectory should be created into which the copied files will be placed.

11 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-11 The two panels below illustrate where the data files are housed in b3 and b4, for example Since the logins are both under cs4cd3, these are local file space configurations. Figure 5: File configuration for the cs4cc3 login on the b3 machine via ssh. Figure 5 and 6 panels illustrate the configuration for the cs4cc3 login filespace, which is shared on rtoo. IN both cases, locate and reproduce in your lab writeup, what the NFS command did to make filespace shared such that it produces the listing under both disk free (df) and ls la commands as below.

12 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-12 Section 2.0 In order to obtain an appreciation for the difference in transfer times when copying files over the network versus copying locally on the same machine perform the following steps. Do this with most of the files present, although do not spend more than about 10 minutes on the largest nor bother with files that require sub-second times to transfer. For each transfer command listed do the same file transfer three times each in a row for each of the different file sizes. Using a script <filename> command may be useful rather than recording results by hand, since there will be so many. Figure 6: File configuration for the cs4cc3 login on the b4 machine via ssh.

13 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page NFS (sitting in front of the x-display server (wolf or fox) logged into cs4cc3 on b3 or b4) time cp /datdir/8m /some_tmp_dir/xyz Do the copy two more times and note how long it takes now and for subsequent copies. Note the time it takes to copy 8m file, 4m and the other files. Record the results and attempt to prognosticate on any consistent differences in timing. Plot the file size versus time required for the cp on NFS filesystems. 2.2 local (sitting in front of wolf or fox and logged into cs4cd3 on b3 or b4) time cp /datdir/4m /some_tmp/xyz time cp [-- second copy] etc. Do the copy two more times and note how long it takes now and for subsequent copies. Note the time it takes to copy 4m file, 2m and the other files. Record the results and attempt to prognosticate on any consistent differences in timing. Plot the file size versus time required for the cp on the local filesystem. 2.3 remote to local (sitting in front of wolf or fox and logged into cs4cd3 on b3 or b4) time rcp b3: 8m b4: 8m where it is assumed the correct subdirectories have been added to the files as prefixes to create the correct file specification path. Do this three times in total for each file and then for the usual set of files after that. In this step also do the reverse operations, that is, b4 to b3 remaining logged into whichever machine that was used in the first part of this section. Compare these times with the rcp that will be done next and comment on any difference found. Why do you suppose there is a difference? 2.3 remote/nfs to local (sitting in front of wolf or fox and logged into cs4cc3 on b3 or b4) time rcp b3: 8m b4: 8m where it is assumed the correct subdirectories have been added to the files as prefixes to create the correct file specification path. Do this three times in total for each file and then for the usual set of files after that.

14 2004/2005 CS 4CC3/6CC3 -- Laboratory 6 page 6-14 In this step also do the reverse operations, that is, b4 to b3 remaining logged into whichever machine that was used in the first part of this section. Is there a consistently different time taken to copy the 8m file from the directory /tmp/cs4cc3a? to the file 8m in the same directory, than to copy the 8m file from your home directory to the /tmp/cs4cc3a? directory? If so explain. In other words, ignoring the effects of buffering, why would remote-to-local be faster than local-to-local or visa-versa? There are several possible reasons -- name as many reasons as you can. Note1: /tmp/cs4cc3a is a directory on the local disk while your home directory is mounted via NFS from a disk on rtoo, as illustrated in figure 3. Note2: to defeat OS buffering of disk blocks when timing copies simply do the following: cp file file.tmp cp file.tmp file This will keep the file the same size and have the same contents but the OS will have to ignore any buffered blocks just in case the file has changed. <End of Part I and Part II> file:4cd04lb6.doc date:24oct04/wfsp revision level: 0