phc-discussions - memory bandwidth usage for hashing vs. other server tasks

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [day] [month] [year] [list]

Date: Fri, 9 Oct 2015 05:52:27 +0300
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: memory bandwidth usage for hashing vs. other server tasks

Hi,

Here's a recent blog post and paper related to a topic that kept coming
up in here once in a while - can we use a lot of memory bandwidth for
password hashing yet not impact other server usage too badly (if the
server isn't dedicated to just password hashing)?

http://danluu.com/intel-cat/
http://csl.stanford.edu/~christos/publications/2015.heracles.isca.pdf

"LLCs for high-end server chips are between 12MB and 30MB, even though
we only need 4MB to get 90% of the performance, and the 90%-ile
utilization of bandwidth is 31%.  This seems like a waste of resources.
We have a lot of resources sitting idle, or not being used effectively.
The good news is that, since we get such low utilization out of the
shared resources on our chips, we should be able to schedule multiple
tasks on one machine without degrading performance.

Great!  What happens when we schedule multiple tasks on one machine?"

A lot of stuff happens, and it takes more than just a quick glance at
the blog post to start to understand it.

Here's my take at it, as it relates to our question above: there's
plenty of idle memory bandwidth on servers even under full load by
memory latency bound tasks, which is typical, however trying to use
this idle memory bandwidth for an extra task (such as password hashing)
might push the other tasks above their latency SLA, even though their
throughput isn't impacted as much.  The paper presents ways to avoid
unacceptable latency impact, but doing so takes effort in a given
deployment (it isn't something that will just happen by default), and
besides password hashing is usually not a best effort task, but is also
an on-demand task.

My opinion remains that we should use computation and memory bandwidth
in a balanced manner, maximizing both at full server load by the
password hashing, yet expecting that in actual usage the server's load
may come from some mix of password hashing and other tasks.  It's
unrealistic to expect that password hashing, whether computation or/and
memory bound, won't have significant impact on other tasks, but it is
realistic to make it such that it degrades other tasks' performance
(both latency and throughput) gracefully as the password hashing request
rate increases.  For example, 50% usage of a server's password hashing
request rate capacity may actually leave 50% of the server's capacity at
other tasks for use by those tasks.

Alexander