MemTest86 and Memtest86+

Christian Külker

0.1.1 2023-03-10

Introduction

There are many ways to test a computer’s memory. Naively, memory can be thought of as a binary object. While in binary systems, such as today’s computers, the state of a given memory cell should be binary; it is not true that the overall state of the device (the memory module) is OK or Not OK (NG), so it can not be considered binary. There are many states between OK and Not OK. While in many circumstances memory appears OK and is considered to be working, it can be seen, not only in supercomputing, that memory as an instance of physics is not binary and there are several levels of OK, depending on usage and user acceptance. For some unknown reason, ECC memory is not considered a mainstream object of use. Therefore, some errors are not detected and therefore accepted by the user, even though the user may not admit it. In supercomputers, memory that throws errors is less welcome. Usually ECC is used where possible, but even then the memory must be validated. Unfortunately, testing and evaluating memory is a difficult and time-consuming business. The reason is that certain value streams written to memory can cause an error while others do not. For example, 11->00->11 might be an error-free pattern, while 11->01->10->11 might not. This is because in certain memory modules, for example, cross-talk between memory lanes is possible, or some other kind of local phenomena.

In an ideal world, a memory test must take into account the physical structure of the memory being tested. While some supercomputers with custom-designed memory spend years evaluating and qualifying memory tests, consumer electronics such as the i386 and successor architectures are typically not tested down to the design level. A more general brute force approach is to write many different patterns to memory and read them back. Doing this structurally and/or randomly over a long period of time will not guarantee error-free memory, but the probability of undetected errors decreases as the error-free time of a test increases.