For many early computers, drum memory formed the main working memory of the computer. It was so common that these computers were often referred to as drum machines. Some drum memories were also used as secondary storage.
Drums were displaced as primary computer memory by magnetic core memory, which offered a better balance of size, speed, cost, reliability and potential for further improvements. Drums in turn were replaced by hard disk drives for secondary storage, which were both less expensive and offered denser storage. The manufacture of drums ceased in the 1970s.
A drum memory contained a large metal cylinder, coated on the outside surface with a ferromagnetic recording material. It could be considered the precursor to the hard disk drive (HDD), but in the form of a drum rather than a flat disk. In most designs, one or more rows of fixed read-write heads ran along the long axis of the drum, one for each track. The drum's controller simply selected the proper head and waited for the data to appear under it as the drum turned (rotational latency). Not all drum units were designed with each track having its own head. Some, such as the English Electric DEUCE drum and the UNIVAC FASTRAND had multiple heads moving a short distance on the drum in contrast to modern HDDs, which have one head per platter surface.
The performance of a drum with one head per track is comparable to that of a disk with one head per track and is determined almost entirely by the rotational latency, whereas in an HDD with moving heads its performance includes a rotational latency delay plus the time to position the head over the desired track (seek time). In the era when drums were used as main working memory, programmers often did optimum programming--the programmer--or the assembler, e.g., Symbolic Optimal Assembly Program (SOAP)--positioned code on the drum in such a way as to reduce the amount of time needed for the next instruction to rotate into place under the head. They did this by timing how long it would take after loading an instruction for the computer to be ready to read the next one, then placing that instruction on the drum so that it would arrive under a head just in time. This method of timing-compensation, called the "skip factor" or "interleaving" (interleaving in disk storage), was used for many years in storage memory controllers.
One of the earliest functioning computers to employ drum memory was the Atanasoff-Berry computer (1942). It stored 3,000 bits; however, it employed capacitance rather than magnetism to store the information. The outer surface of the drum was lined with electrical contacts leading to capacitors contained within.
Magnetic drums were developed for the U.S. Navy during World War II with the work continuing at Engineering Research Associates (ERA) in 1946 and 1947. An experimental ERA study was completed and reported to the Navy on June 19, 1947. Other early drum storage device development occurred at Birkbeck College (University of London), Harvard University, IBM and the University of Manchester. An ERA drum was the internal memory for the ATLAS-I computer delivered to the U.S. Navy in October 1950. Through mergers, ERA became a division of UNIVAC shipping the Series 1100 drum as a part of the UNIVAC File Computer in 1956; each drum stored 180,000 characters.
The first mass-produced computer, the IBM 650, had about 8.5 kilobytes of drum memory (later doubled to about 17 kilobytes in the Model 4). As late as 1980, PDP-11/45 machines using magnetic core main memory and drums for swapping were still in use at many of the original UNIX sites.
In BSD Unix and its descendants, /dev/drum was the name of the default virtual memory (swap) device, deriving from the use of drum secondary-storage devices as backup storage for pages in virtual memory.