[GSOC] System V IPC in userland

[Grigore Larisa]

Hi Samuel,

On Mon, Sep 30, 2013 at 12:45 AM, Samuel J. Greear <sjg at evilcode.net> wrote:

> Larisa,
>
> Could you further describe these attached results and what exactly is
> being tested / how it is being tested? (Sorry if you had done this in a
> prior email and I just missed it).
>
> Thanks for the great work this summer!
>
> Sam
>


I have run pgbench, in the first scenarion using kernel implementation
(without libc wrappers) and, in the second scenario, using userland
implementation (with libc wrappers and setting environment variable USR_SYSVIPC
to "1"). I've used parameters from the first three columns (number of
clients, number of threads and duration) for pgbench and I've collected the
number of transactions made including connections established (column I.)
and excluding connections established (column E.).

I've noticed there is a problem with column I. for the userland
implementation (there is a big difference between kernel and userland
implementation) and I don't understand why it is not the same behavior for
column E. Maybe this will help me understand where it can be done anything
for performance improvements.

The benchmarks are made on a KVM with 4 cores and 4G RAM that runs on a
computer with 4 cores HT (8 threads).

Larisa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dragonflybsd.org/pipermail/kernel/attachments/20131001/94bececd/attachment.html>