Re: NSE Nsock Binding Performance Results and Improvements

[Patrick Donnelly <batrick () batbytes com>]

On Wed, Jul 15, 2009 at 9:39 AM, Patrick Donnelly<batrick () batbytes com> wrote:
> On Mon, Jul 13, 2009 at 11:58 PM, Patrick Donnelly<batrick () batbytes com> wrote:
> > Attached is a PDF of the report.
>
> Here is a summary of the report:
>
> We conducted tests for three different script sets across two machines
> with manually set parallelism values ranging from 10 to 100 (the
> current default parallelism being 10). Graphs of numerous concurrently
> running script threads (up to 900) consistently formed an exponential
> decay regression for script scan time. One running script
> (html-title.nse) would usually see a constant flat improvement across
> the range of parallelisms (that is, no change in total scan time). The
> timings for the scans showed that a parallelism of 40 got the most
> benefit for each increase in parallelism.
>
> Due to our results, we changed the parallelism value to 40 to get
> increased performance by default out of NSE (revision 14264).

After recent discussion we have again changed the parallelism to a
lower value of 20 (revision 14303). The reasoning for only doubling
the previous parallelism of 10 is concerns about a significant niche
of NSE users that will be running on poor connections incapable of the
load demand a parallelism of 40 may introduce (especially at the
start). It should be noted that a parallelism of 20 is still a huge
improvement over 10 and in some scans caused a scan to take less than
half the time to run.

The current plan is to investigate a more dynamic solution that will
allocate the optimal amount of open sockets for the scan based on real
time network conditions.

-- 
-Patrick Donnelly

"Let all men know thee, but no man know thee thoroughly: Men freely
ford that see the shallows."

- Benjamin Franklin

_______________________________________________
Sent through the nmap-dev mailing list
http://cgi.insecure.org/mailman/listinfo/nmap-dev
Archived at http://SecLists.Org