pantheon of distributed operating systems

Konstantinos Foutzopoulos

At the mid-80s to early-90s there was much research going in the development of distributed operating systems (see lecture and paper by Tanenbaum). Those are loosely characterized by their purpose to provide a single system to a collection (network) of workstations that connects (not necessarily joins) them and makes them act like one. For many such systems, distributed implies transparency. Explaining, the system should blur location of the resources. That is whether a resource is local or remote should be irrelevant.

Though now mostly lost in history, they still hold a lasting influence and interest as those problems are still relevant. The most successful of them, taking into account academic publications, commercialization, and popularity, (and personal opinion) were Sprite, Amoeba, and Plan 9. The last one especially continues to be popular among those interested in operating systems.


Amoeba (see intro paper and status report) was developed at the Vrije Universiteit Amsterdam by Andrew Tanenbaum’s research group between 1981 (initial work; first proto release in 1983) and 1996. It was commercial but freely available for academic usage. Amoeba was intended for both distributed and parallel computing. Though a mechanism for doing both applications was provided, the policy was determined by user-level programs.

In order to achieve performance over the network, it used the high performant FLIP, short for Fast Local Internet Protocol. This protocol allowed for clean, simple and efficient communication between distributed nodes. Machines that had more one network interface automatically act as FLIP router between the network and therefore connected various LANs together.

Amoeba was a system written from scratch. Though it had a Unix-like interface, the kernel was designed with a microkernel architecture. This makes it among the first, and basically the very few even now, successful (in that it was used for day-to-day computing) microkernel operating systems. That architecture means that the system has a modular structure and is a collection of independent processes. Those processes can either be application programs, called clients, or servers such as drivers. The microkernel basic function is to provide an environment inside which client and servers communicate. Its job is to support threads, remote procedural call (RPC), memory management, and I/O. Everything else is built atop those primitives.

Some Amoeba-specific software outside the kernel is:

  • Boot server, that controls all global system servers outside the kernel.

  • Bullet file server (akin to filesystem), that stores files contiguously on disk and caches, giving it high speed, whole files contiguously in core. A user program needing a file will request that Bullet sends the entire file in a single RPC. As Bullet uses the virtual disk server to perform I/O to disk it’s possible to run it as a normal user program.

  • SOAP directory server, that takes care of file management and naming. The file and directory server functionality in traditional (monolithic) operating system are part of the kernel.

Also closely related to Amoeba (the teams cooperated intensively) was Orca, a programming language developed specifically for parallel programming. Orca allows creation of user-defined data types which processes on different machines can share in a controlled way. This in effect simulates an object-based distributed shared memory over a network.

A typical Amoeba installation will have dedicated terminals, a processor pool, and specialized servers (file and directory). Those can be run on processor pool or dedicated hardware. It may be seen that this is just mainframe (time-sharing) computing, but note that processor pool doesn’t mean a single computer or that access is granted for only specifically requested taks. Rather the processor pool, that can as well be a computer cluster, from user’s point of view is part of the system.

Years since v5.3 was released in 1996, an unofficial revised distribution called FSD was released and was kept in development until it was frozen in 2002. FSD includes many kernel extensions and changes, new hardware driver-related addition, and improvements for third-party software. The system was used for developer’s physical studies. It can be downloaded from its SF project page.

Ending this section, Guido van Rossum created Python for this operating system. Amoeba has its own system call interface that wasn’t easily available from shell. Alongside the lack of exceptions in shell lead Guide to write a generally extensible language with ABC-like syntax, a language whose implementation Guido had experience with, and exceptions inspired from Modula-3.


Sprite was developed at the University of California by John Ousterhout’s research group between 1984 (initial work; 1987 was used for day-to-day computing) and 1992. It was Unix-like but was written from scratch. It was used as testbed for research in various topics, among them network, log-structured, and striped filesystems.

At the time (updated) the project started there were no good network filesystems and even predated NFS. Also, administration of a network of workstations was difficult. Its goal was to build a new operating system designed for network support from the start. In the end it had a few technical accomplishments that stand out.

  • Sprite’s network filesystem utilizing file caching (see paper) for high performance, which allowed file sharing transparently between workstations. By implementing I/O using pseudo (in modern terms, virtual) devices and filesystems (see paper), it allowed access to I/O devices uniformly across the network.

  • Process migration mechanism (see paper), which allowed processes sharing (that is processes could be moved) transparently between workstations. The mechanism kept track of idle machines and evicted migrated processes when local resources were again required (when the workstation user returned). Ideas from process migration have found their way in virtual machine monitors (hypervisors).

  • Single-system image, which means Sprite appears and feels like a single system with storage and processing resources shared uniformly among the workstations. Administration was scale invariant. Adding a new machine was not any different than adding a new user account. Sprite also supported different machine architectures in the same cluster by utilizing a framework that separated architecture-independent and architecture-specific information. This model was adapted for cluster and grid computing in the form of cluster management systems. An example is MOSIX, that provides single-system image capabilities to Linux.

  • Log-structured filesystem (see paper), which was a new approach to filesystem design. This kind of system treated the disk like a tap. By having information written sequentially it allows better efficiency especially when reading and writing small files. The last one resulted by having many large sequential writes batched together. Other advantages were fast crash recovery and block size variation per file. Among the most advanced log-structured filesystem nowadays are NOVA, developed at the UoC same as Sprite, and F2FS, developed by Samsung. F2FS is especially designed for flash memory-based storage devices and has been adopted by few Android devices.

  • Zebra (see paper on design and on implementation), a distributed file system that increased throughput by striping files across multiple servers giving it scalable performance. Zebra also wrote parity information for each stripe similar to RAID arrays. This could allow system operation to continue even in the event that a server is unavailable giving it high availability. The striping approach, in contrast to the file-based approach of other stripping filesystems, utilized techniques from log-structured filesystem. This simplified parity mechanism, reduced parity overhead, and allowed clients to batch together small writes giving it high server efficiency.

A paper written by Douglies, Ousterhout, Kaashoek, and Thnenbaum, compares Amoeba and Sprite. Although they’ve similar goals, they diverge philosophically. Specifically on whether distributing computing or traditional Unix-style applications should emphasized, and on whether a combination of terminal and shared processors or a workstation-centered model (workstations and file servers) should be used. Those diverged philosophies result in the following differences.

Amoeba Sprite
object-based shared filesystem-oriented
processor pool process migration model
server-only caching client-level caching available
microkernel monolithic
(hence) shared services (hence) services redunancy
user-level IPC mechanism pseudo-devices and files

Microkernel approaches for Sprite were also explored. Specifically Sprite kernel was ported to run as user-level server process on [Mach microkernel], which was developed at Carnagie Mellon University between 1985 and 1994. The Sprite server was smaller than the original kernel and it contained almost no machine specific code. As a downside it was benchmarked at about 1/3 of the native perfomance. An expected issue as bad performance was common to microkernels of the era.

Sprite’s demise

The project ended when the kernel became hard to maintain for its small development team. Another problem was the inability to catch up with features added on commercial UNIX systems of the era. Those features weren’t research oriented making those tasks mundane for a foremost research operating system.

The source code is archived on OSPreservProject. A precompiled DECstation image is also available that can run using GXemul. Downloading the image and starting the emulator is done with (adapted from GXemul docs):

wget https://github.com/OSPreservProject/sprite/raw/master/ds5000.bt
gxemul -X -e 3max -M128 -d ds5000.bt -j vmsprite -o ''

At the first boot up the following network settings should be entered.

Your machine's Ethernet address:    10:20:30:00:00:10
Your machine's IP:        
Subnet mask:                        0xff000000
Gateway's Ethernet address:         60:50:40:30:20:10
Gateway's IP:             

Note that the bootable Sprite image is merely a demonstration, rather a robust system. It misses floating point and network support. Once logged in someone can run xinit to start the X11 environment.

Ending this section, John Ousterhout also created the Tcl scripting language and the Tk widget toolkit alongside this operating system.

Plan 9

Plan 9 from Bell Labs (see intro paper) originated at Bell Labs developed by the fathers of Unix between late-80s (first public release in 1992) and 2002. It has been refered to as what Unix should have been. Initially commercial but freely available for academic use, after its demise it became source-available (license didn’t qualify as open source, let alone free software) and few years ago was re-licensed under GPLv2.

Everything starts out with 9P, a lightweight network-level transport-independent protocol that enables transparent access to resources independent of their location. All the resources are represented as files within an hierarchical filesystem. Having resources named and accessed like files in an hierarchical filesystem implies the everything is a file upon which Unix was designed is taken to extreme. On top, a naming system is built that lets processes build customized views of the network resources. Therefore, Plan 9 lets a process build a private environment rather doing all computing on a single workstation. This model of per-process namespaces and filesystem-like resources is then extended throughout the system.

An example where this is applied, is rio (and 8½ before it), its windowing system, that is implemented as a file server. Each window is created in a separate namespace. Thanks to its 9P interface, rio is network transparent even without including any network-aware code. This is in contrast to X server that specifically includes such code. Also in contrast to X server, a remote rio application sees its files in the usual location and therefore doesn’t know whether it runs locally or remotely.

It should be noted that X used in previous systems bears an extra complexity not required as, even without its network-capabilities, windows can be run over the network thanks to the distributed nature of those systems.

In a way it combines the approach of the previous two systems. Similar to Sprite, it is a monolithic system and utilizes virtual filesystems, whereas similar to Amoeba, considers that computers can handle different tasks and the importance of parallel computing. Expanding on the later, A typical Plan 9 installation will have dedicated terminals, CPU servers, and file servers. The CPU server differs from Amoeba processor pool. Plan 9, rather making use of many processors, it uses a few multiprocessors. This is also the reason a process migration mechanism isn’t provided. Note that a computer can be set up for any task. Someone can even have all tasks handled in the same system.

Historically, used in first and second edition, Alef was developed, a concurrent programming language with C-like syntax and semantics. It holds heritage from Newsqueak, another language developed previously at Bell Labs. Similar to it, Alef had a CSP (channel-based) concurrency style. In order to maintain one less language, it was abandoned in favor of thread library for C. Plan 9 favoring concurrency over parallelism relates to the philosophical difference with Amoeba (see paragraph on Orca) on how processors should be handled.

In contrast to Amoeba and Sprite that were specifically designed for local-area networks, Plan 9 considers the existence of wide-area networks and grids could exist even over low speed connections (internet). It is assumed that a high speed connection exists between CPU and file servers though it isn’t a hard requirement. The environment can be configured to be as distributed or centralized as required by its usage.

The latest version can be downloaded from the mirror of the official site. Virtual disk images can be found in 9legacy site. This site also offers numerous patches that can be applied to the latest Plan 9 release. Another option is 9front, an actively developed fork. Also checkout Plan B, a set of user programs that run on top of Plan 9.


Plan 9 was to be succeeded by Inferno (see intro paper). Also started by Bell Labs, it is now developed and maintained by Vita Nuova. Inferno is based on Plan 9 ideas but takes a more radical approach, as seen by the following new aspects.

  • Limbo, a new garbage-collected concurrent language with C-like syntax. It can be considered a director successor to Alef. The very popular nowadays Go contains ideas found in Limbo.

  • Dis, a virtual machine designed for portability and JIT.

  • Runs on hardware standalone or hosted. Explainining, Inferno can run as a user application on top of an existing operating system.

This design (a virtual machine) that offers the ability to run Limbo-written software atop another system bears similarity to Java’s initial goal to write once run anywhere. Though the systems have similar goals, the implementation differs.

Plan 9 in Cyberspace

Both Plan 9 and Inferno had their development stopped because they brought no financial gains enough to justify continued investment in Bell Labs part. Today, and basically this article was written because of it, Bell Labs (now owned by Nokia) announced that is transferring copyright to Plan 9 Foundation incorporated by the Plan 9 creators allowing development to carry on. Along with this, all Plan 9 releases were relicensed under MIT.

Concluding this section, it should be noted that Plan 9 is known for far more than its distributed model, and posts covering those other aspects should eventually follow.