block level: new tool "blkreplay" for tests / benchmarking

block level: new tool "blkreplay" for tests / benchmarking

[Thomas Schöbel-Theuer]

Hi all,

in some sense, blkreplay is just the opposite of the well-known kernel
tool blktrace: recordings of (natural) loads can be replayed in the
laboratory, measuring IO latencies (and other indicators) for later
visualization (so-called "sonar diagrams"), e.g. for comparison of IO
schedulers, or driver overhead, or network latencies, or hardware.

While blkreplay itself is a rather simple tool, the real value of the
GPL'ed blkreplay project might be the large collection of natural loads,
stemming from datacenters at 1&1.

The public collection of natural loads covers a broad range of
applications, such as databases (even commercial ones like Oracle),
shared hosting, virtualized servers, etc. Some loads are even 24h
recordings. The collection will be enhanced during the next few days
(internally at 1&1, we have already more than 1000 blktrace recordings,
so our main problem is selecting the "interesting" ones, otherwise you
could be overwhelmed).

blkreplay comes with a comprehensive manual of already 57 pages, which
will be extended during the next weeks.

Introduction / Abstract: www.blkreplay.org or
https://github.com/schoebel/blkreplay/wiki

Manual:
https://github.com/schoebel/blkreplay/raw/master/doc/blkreplay.pdf

Source code: https://github.com/schoebel/blkreplay

Collection of loads: http://www.blkreplay.org/loads/

Enjoy,

Thomas