memtester is a utility for testing the memory subsystem in a computer to determine if it is faulty. The original source was by Simon Kirby [email protected]. Charles Cazabon has by this time completely rewritten the original source, and added many additional tests to help catch borderline memory. Charles Cazabon also rewrote the original tests (which catch mainly memory bits which are stuck permanently high or low) so that they run approximately an order of magnitude faster.

The version 4 rewrite was mainly to accomplish three things:

1. the previous code was basically a hack, and was ugly.
2. to make the code more portable. The previous version required some hackery to compile on some systems.
3. to make the code fully 64-bit aware. The previous version worked on 64-bit systems, but did not fully stress the memory subsystems on them -- this version should be better at stress-testing 64-bit systems.

memtester uses CMake as build tool. The basic build can be done as follows:

mkdir build
cd build
cmake ..
make
#optional
sudo make install 

Memtester 4 has been sucessfully build on the following systems:

• HP Tru64 Unix 4.0g (Alpha)
• HP Tru64 Unix 5.1b (Alpha)
• HP-UX 11i 11.11 (PA-RISC)
• HP-UX 11i 11.23 (64-bit Itanium)
• Debian GNU/Linux 3.0 (various)
• other 32-bit Linux (RedHat, SuSE, Ubuntu, etc) (various)
• RedHat Enterprise Linux/CentOS (64-bit AMD Opteron)
• FreeBSD 4.9 (32-bit Intel)
• FreeBSD 5.1 (64-bit Alpha)
• NetBSD 1.6 (32-bit Intel)
• Darwin (OS X) 7.5.0 (32-bit PowerPC)
• OS X Leopard/Panther/whatever -- 32- or 64-bit, PPC or x86

It should, however, work on other Unix-like systems. If you have trouble building memtester on your system, please report it to me so it can be fixed.

Usage is simple for the basic case. As root, run the resulting memtester binary with the following commandline:

memtester <memory> [runs]

where is the amount of memory to test, in megabytes by default. You can optionally include a suffix of B, K, M, or G (for bytes, kilobytes, megabytes, and gigabytes respectively). [runs] is an optional limit to the number of runs through all tests.

An optional "-p physaddr" argument available to cause memtester to test memory starting at a specific physical memory address (by mmap(2)ing a device file representing physical memory (/dev/mem by default, but can be specified with the "-d device" option) starting at an offset of physaddr, which is given in hex).

Note: the memory specified will be overwritten during testing; you therefore cannot specify a region belonging to the kernel or other applications without causing the other process or entire system to crash). If you use this option, it is up to you to ensure the specified memory is safe to overwrite. That makes this option mostly of use for testing memory-mapped I/O devices and similar. Thanks to Allon Stern for the idea behind this feature. For example, if you want to test a bank of RAM or device which is 64kbytes in size and starts at physical address 0x0C0000 through the normal /dev/mem, you would run memtester as follows:

memtester -p 0x0c0000 64k [runs]

If instead that device presented its memory as /dev/foodev at offset 0, you would run memtester instead as follows:

memtester -p 0 -d /dev/foodev 64k [runs]

Note that the "-d" option can only be specified in combination with "-p".

memtester must run as user root so that it can lock its pages into memory. If memtester fails to lock its pages, it will issue a warning and continue regardless. Testing without the memory being locked is generally very slow and not particularly accurate, as you'll end up testing the same memory over and over as the system swaps the larger region.

The following tests are from the original version, updated simply for speed and rewritten to fit the new framework of the program. These tests will mainly catch memory errors due to bad bits which are permanently stuck high or low:

• Random value
• XOR comparison
• SUB comparison
• MUL comparison
• DIV comparison
• OR comparison
• AND comparison

The following tests were implemented by Charles Cazabon, and will do a slightly better job of catching flaky bits, which may or may not hold a true value:

• Sequential Increment
• Block Sequential
• Solid Bits

The remaining tests were also implemented by Charles Cazabon, and are designed to catch bad bits which are dependent on the current values of surrounding bits in either the same word32, or in the preceding and succeeding word32s.

• Bit Flip
• Checkerboard
• Walking Ones
• Walking Zeroes
• Bit Spread

There is also a test (Stuck Address) which is run first. It determines if the memory locations the program attempts to access are addressed properly or not.
If this test reports errors, there is almost certainly a problem somewhere in the memory subsystem. Results from the rest of the tests cannot be considered accurate if this test fails:

• Stuck Address

Please report bugs by opening an issue or submitting a PR on the Github project.

As mentioned, memtester v4 was tested with a wide variety of Unix- and Unix-like systems. However, at least two issues with ancient HP/UX versions, and one with Tru64, have come up since.

HP/UX versions prior to v11 do not support mlock() and will fail with an invalid syscall error at runtime. If you're building on HP/UX v10.20 or similar, change int do_mlock = 1 around line 128 of memtester.c to int do_mlock = 0 and rebuild. They also are missing strtoull() from C99. To fix this build error, add #define strtoull __strtoull to the top of memtester.c; this should at least work with a GCC toolchain.

Tru64 is missing strtoull() from C99. To fix this build error when using cc, add #define strtoull strtoul to the top of memtester.c; this is safe because long on Tru64 is 64-bits. This build problem could be dealt with automatically, but I don't particularly want to invest the necessary time, because Tru64 went out of support in 2012. In addition, I no longer have access to a build machine for testing on this platform.

by Charles Cazabon [email protected]

forked in 2024 by Benedikt Bergenthal [email protected]

Copyright 1999 Simon Kirby.

Version 2 Copyright 1999 Charles Cazabon.

Version 3 not publicly released.

Version 4 rewrite:

Copyright 2004-2020 Charles Cazabon.

2024 Benedikt Bergenthal [email protected]

Licensed under the terms of the GNU General Public License version 2 (only).

See the file COPYING for details.