By subject
A subject. The loosest grouping, and the one to reach for last.
One line, 1948 to 2020: a device is invented, a theory makes it manufacturable, a scaling rule makes it shrinkable, an extrapolation makes the shrinking an industry expectation, and then the physics stops cooperating. Read in order the argument is legible. Bardeen and Brattain announce the point-contact transistor in two pages. Shockley supplies the junction theory that makes a manufacturable device possible. Dennard states the rule by which a smaller transistor is also a faster and cooler one, which is what made shrinking free. Moore, in a trade magazine and from five data points, extrapolates the component count ten years out; the industry then organises its roadmaps around the extrapolation and it becomes a law by being treated as one. Bohr, thirty years on, asks what Dennard's paper actually said. Theis and Philip, and Shalf, describe what happens when Dennard scaling ends but Moore's curve is still expected. The through line is that each step buys the next one's assumption, and the last two records are about the bill.
The Computer Systems course is described as operating systems, performance versus information security. This is that argument as a line of primary sources, and it has two halves fifty years apart. Lampson states the problem in 1973: a confined program must not leak, and the channels by which it can are more numerous than anyone enumerates — he names the covert channel. Saltzer and Schroeder give the design principles meant to hold the line. Then a long gap in which the principles are taught and the hardware quietly acquires caches, branch predictors and speculation, each a performance feature and each a shared resource. Ge, Yarom, Cock and Heiser catalogue the resulting timing channels in 2016, a year before anyone showed how bad it was. Meltdown and Spectre then show it: Meltdown an implementation defect that a software mitigation closes at a bounded cost, Spectre a consequence of speculation itself that does not close. Klein and colleagues' seL4 sits deliberately in this set as the other side of the boundary — a kernel proved to implement its specification, in a corpus where the attacks all live beneath the specification. The through line is that isolation is asserted at one level and broken at a lower one, and that every performance feature which shares state between principals is a candidate for the next break.