John Fletcher

by John Fletcher

An early feature (preceded only by remote interactive terminal access) added to the Octopus network at LLNL, the Lawrence Livermore National Laboratory, was archival storage. The idea was to provide rapid access by the large mainframe computers to the great volumes of data produced by physics codes. A typical one of these codes could run for many hours on CDC 6600s and 7600s or on Crays. Without centralized archival storage, these codes had to rely on magnetic tapes and suffer from the delays involved in fetching, mounting, and dismounting them.

We organized the central filing system around the then new idea of directories. We followed closely what we understood to be the original concept devised by the architects of Multics. Our approach differed from what is common today in at least two significant ways:
  1. Our directory structure was a general directed graph that was intended, eventually, to include, not just the archival files, but all the files in the network. So, entries pointing to a single file or directory could occur in two or more directories. All these entries were of equal status, in contrast with systems in which one entry is the "real" one and the others are "aliases", "shortcuts", or whatever. A result was that closed loops could exist: starting at a directory one could follow entries through a sequence of directories until, after one or more steps, the starting point was reached again. This generality, of course, presented certain difficulties not present in hierarchical systems (such as when exhaustively searching a substructure), but it also gave users a flexibility that often proved very useful. However, the primary consideration in permitting generality was that such was difficult to preclude in a truly distributed directory system: Each of two or more computers might want to make a directory substructure kept on one of the others be a substructure of its own. It therefore would require considerable "handshaking" to preserve a hierarchy, and some of the computers would, through no fault of their own, be unable to create entries that their users thought would be useful.

  2. Each user had a "root" directory, and any directory or file that could be reached from that directory was accessible by him. (Access through some entries, however, could be limited, in particular to read-only.) So, very precise file sharing was possible: Any user could give any other user entries pointing to exactly those files intended for sharing. This may be contrasted with systems in which sharing is limited to being either with everyone else or with exactly those belonging to certain administratively-defined "groups".
Our first really good archival storage device arrived around 1970 and remained in use for over a decade. It was the IBM Photodigital Store or, as we called it, the Photostore. It was used for files, not for directories (which were kept on more conventional devices). The storage medium was high-density silver halide photographic film. The device itself was an electrical, mechanical, and chemical marvel. When I first saw it under construction at the IBM facility in San Jose, I had serious doubts that such a Rube Goldberg device would ever work, but indeed it did. The life history of a particular "chip" of film (about the size of a small playing card) was as follows:
  1. Thiry-two unexposed chips arrived at the Laboratory in a small plastic box or "cell", which was a little smaller than a pack of cigarettes. These were packed in a carton of ten, wrapped in a black cover to exclude light. When more "raw" film was needed, an operator would open a hatch in the top of a low section of the machine and shove the carton in, end-first. Blades would rip the box open, and the cells would drop into a queue from which they would, one by one, advance to the next step: exposure and development.

  2. When a cell reached the head of the queue, its lid would be removed by depressing a release catch, and the chips would be mechanically extracted, one at a time, and held in the beam from an electron gun. The electron beam was magnetically aimed so as to encode the stream of data to be written, forming it into a sequence of dark and light spots on the chip. Between chips, the magnetic field would be sensed and adjusted so as to assure that it was focussed precisely enough to create the tiny spots that were needed. At specified intervals the filiments of the gun would be changed automatically by rotating a turret of eight filaments; only after these were exhausted would operator intervention be required.

  3. Next, a mechanical picker arm would move the chip to a beaker on a carrousel. At successive stations, as the carrousel rotated, the chip would be exposed to developer, stop, fix, wash water, and dry air. Then a picker arm would move the chip into a waiting empty cell. This cell was actually the cell that had just previously been at the head of the queue for the electron gun (so the machine had to be "primed" with one empty cell). When the cell was full of 32 chips it would be "sucked" by a vacuum through pneumatic tubes to its resting place in an "egg crate" that held 30 cells. These crates were stacked in several wings of the machine. The total storage capacity of the machine was one trillion bits (10**12 bits), far more than was to be found in other storage devices of that time.

  4. When a chip was to read, it was pneumatically extracted from its crate and sucked to a reading station, where a picker arm held it in front of a flying spot scanner that read the spots that had been previously written. The recorded bits included an error correcting code that enabled the reader to proceed even in the face of flaws in the film. (A newly recorded chip was always given an immediate test reading to make sure that the recording was readable; if not, it was rewritten.)

  5. As should be clear from the above description, the Photostore was a write-once device. Our filing system recognized this explicitly by declaring all files on the device to be read-only after their initial recording (which was part of the file-creation process). We could, of course, have provided the illusion of a read-write medium by a system of rerecordings, but such would be very wasteful of film. In the interest of efficiency, we wanted the Photostore to look like what it really was: an archive and nothing else. In fact, the chips lasted "forever". When all the crates were full, the least-used cells were transferred to other crates designed to be stored offline.
However, "forever" was not really forever. Eventually, the Photostore became obsolescent, and IBM told us that they would stop making replacement parts. Many at the Laboratory did not want to give up the device and urged that we somehow maintain it on our own. However, a succession of study committees, each convened by those unhappy with the decision of the preceding committee, all concluded that there was no practical way to keep the machine going. So, one by one, all the chips, even those that had been taken offline, were called back to the reader for one last time, and their data was transferred to a more modern device.

For more technical information on the Photostore System, see the following:
The Photostore Principles of Operation Manual

Images of the Datacell and the Photostore
Figure 1: The Photostore Cabinets
Figure 2: The IBM 2321 Data Cell
This device, the IBM 2321 Data Cell, is shown here mostly to allow comparison with the 1360. The Data Cell had 200 strips of wide magnetic tape each about 40 cm long, and there were 10 Data Cells per storage unit. The total capacity was about 4 x 10^9 bits. A detailed description of the 2321's operation and speeds is beyond the scope of this picture caption. Suffice it to say that it was slower and accommodated many fewer bits (10^9) than the Photostore (10^12).
Figure 3: Another View of the Photostore
Figure 4: Photostore Chip Storage Unit
An IBM representative and Dr. Sidney Fernbach, Head of the Computation Department, are standing by the opened Photostore chip storage unit - we had two such units. Each held almost 4,000 film chip boxes. They are examining the film used in the chip boxes. This view also shows the crate-like box used to insert boxes of fresh film or previously written chips into the 1360. Each chip box held 32 film chips and each film chin provided 32 data fields. Above the crate-like box one can see 4 trays, each holding about 1,000 chip boxes.
Figure 5: A Chip Box Tray
John Fletcher is holding a tray used to enter fresh film into the 1360, and a film chip box.
Figure 6: Cartridge Storage Unit
An unknown woman is standing in front of the Cartridge Storage Unit (one of two) of the IBM 1360, the Photodigital Store. The box she is holding held 32 chips of film, with each chip containing data in a 4 x 8 array of 32 data fields. The tray was used to enter or remove chip boxes. There were about 7,700 chip boxes, for a total capacity of 10^12 user data bits. In addition, there were about a 30% extra bits for error control.
Figure 7: A Diagram of the Photostore
Figure 8: A Schematic of the Photostore
Figure 9: A Schematic of the Photostore Reader
Figure 10: A Film Chip Box
The film chip box is shown with the top cover removed. The total size of the box is approximately 3 x 1.1 x 1.6 inches. Each film chip held 32 data fields in a 4 x 8 array. One box was very roughly equivalent to one 2400 foot, 8 channel, magnetic tape recorded at 800 bits per inch.
Figure 11: The Photostore Electron Beam Writer and Developer
Figure 12: A Film Chip
The 32 data fields on the chip are clearly visible. The chip is approximately 2.75 x 1.38 inches, and had a capacity of ~6.6 x 10^6 total bits, of which there were ~4.7 x 10^6 user data bits.
Figure 13: Film Chip Data Fields
This is an enlargement of a portion of a film chip. The shadowy object is the eye of a small needle. The tracks are written and read boustrophedonically (look it up).
  Figure 14: Film Chip Data
A much enlarged portion of one data field. Starting with a track of about 4.2 microns width, each data bit is a pair of areas of 8.5 x 8 microns. One of the areas is left clear, and the other is made opaque by the electron beam, modulating a 17.5 megahertz sine wave. For example, a zero bit is a dark area followed by a clear area. And of course, a one bit is just the opposite. As noted in photostore 5, reading and writing is carried out boustrophidonically, and each line carried approximately 30% extra bits for addressing and special error checking and correction. It is no exaggeration to note that IBM's automatic error control and correction was absolutely vital to the successful use of the 1360.

For more information about John Fletcher, see his two interviews (1, 2) published on this history site.