As a contributor to aolserver and naviserver i want to add a
few comments
- we are running between 30 and 50 servers for various
projects,
i would say that 70% are naviserver right now.
- the reason we switched from aolserver to naviserver was that
with our load-pattern (using OpenACS) we experienced some
problems:
* up to about 1000 concurrent users aolserver was
perfectly fine
* above this, we saw crashes, running out of resources
(connection threads),
memory growth, etc., thread lockups, micro-lockups for
a few seconds.
Some of these lead to contributions to aolserver i did
in the past
* to pinpoint the problem we moved to Zoran's setup
(tcl version, naviserver), that went though heaving
testing on his side
and was rock-stable
* some of our problems disappeared/changed, some not
(burst creation
of theads,...). We have quite different load patterns
than zoran.
* we found sources for our problems at various places
(server, tcl, ...
machine architecture) depending as well on e.g. tcl
versions etc.
By now, most of the problems are gone, we are using
NaviServer in
production since more than two years. A summary is on the
link referenced below. Even more recently, we exchanged the
hardware to a more mainstream one (this improved the
performance by a factor of 3-4!). The fact that e.g. the
resource
consumption went down, helped a lot to run on a much cheaper
machine
(memory consumption, max number of connection threads went
from 80 to 30, etc.).
Btw., in this process of moving from POWER to Intel apparently
the biggest source of our crashes went away. The way Tcl
handles
thread-local storage (Tcl 8.5) seems to cause under heavy load
race conditions, which lead to crashes in otherwise stable
code-pieces (e.g. regexp). I rewrote some of the usages of the
tls infrastructure in tcl to use GCC's non-standard tls
handling via
__thread, then the problem went from regexp to other
places using tls). The problem was most likely dirty reads
in tcl +
mutex handling + POWER + gcc (from rhel). Tcl 8.6 is
supposed to be better in this regard.
For the changes in naviserver, see [1]. With the recent
versions of
naviserver/tcl 8.5/libthread the server runs in a stable
memory size, without
the need for daily reboots (although a reboot has some nice
self-healing
properties for nsvs, etc.). See [2] for a statistics from a
machine with
two naviservers with different configuration running (alice,
nm).
Among other things one can see the stable rss and vsizes of
the servers
over the last few months.
Aolserver is in terms of memory leaks not so bad either. One
can see
on [3] the statistics from openacs.org and
translate.openacs.org which
is runing aolserver 4.5.1. One can see, where we fixed an
application
leak in May [4].
[1]
https://bitbucket.org/naviserver/naviserver/src/5df3b1cb9ea6/NEWS
[2]
http://alice.wu.ac.at/munin/wu.ac.at/alice.wu.ac.at/index.html#naviserver
[3]
http://openacs.org/munin/localdomain/localhost.localdomain/index.html#naviserver
[4]
http://openacs.org/munin/localdomain/localhost.localdomain/naviserver_translate_memsize.html
Concerning the comments below
- the documentation of naviserver is at least par with
aolserver
(the man pages are quite good).
- for me, the the biggest pain is the aolserver-naviserver
config file conversion, but the actual documented
config files on bitbucket contain now all values read
from the
server with the default values.
- porting all the changes from naviserver into aolserver is
much more work than the other way round. i have no
problem with the coexistence of naviserver and aolserver,
providing urgent changes to both servers (as i have done
in the past).
- both aolserver and naviserver are stable and mature (having
advantages and disadvantages), the people running large sites
are rather conservative. Having alternatives is rather a
selling
argument. If e.g. aolserver is dropping windows support,
naviserver can continue it (or vice versa).
-gustaf neumann
On 27.09.12 23:25, John Buckman from BookMooch wrote:
Naviserver has added a lot of interesting features, and appears to be fairly
mature.
I would have probably switched to Naviserver two years ago if they had
documented some of their changes. The quantity of the contributions, and the
interesting nature of many of them, make me feel that Naviserver is far from
end of life.
When I switched (temporarily) to naviserver I found enough things that didn't
work like aolserver, yet were totally undocumented, that the experience was
very frustrating and I went back to aolserver. I was spending too much time
reading C source code to figure things out.
So... my personal vote for an aolserver v5 would be merging in lots of the
naviserver code changes into aolserver. There's a lot of bang-for-our-buck
there. Or, simply running with naviserver, if we (the aolserver community)
can get it
