Thanks for this write up Thierry,
On 19 Oct 2018, at 17:15, thierry bordaz <tbordaz(a)redhat.com>
wrote:
Hi,
C10K is a scalability problem that a server can face when dealing with events of
thousands of connections (i.e. clients) at the same time. Events can be new connections,
new operations on the established connections, closure of connection (from client or
server)
For 389-ds, C10K problem was resolved with a new framework Nunc-Stans [1]. Nunc-stans was
first enabled in RHDS 7.4 and improved/fixed in 7.5. Robustness issues [2] and [3] were
reported in 7.5 and it was decided to disable Nunc-stans. It is not known if those issues
exist or not in 7.4.
William posted a PR to fix those two issues [4]. Nunc-stans is a complex framework, with
its own dynamic. Review of this PR is not easy and even a careful review may not guaranty
it will fix [2] and [3] and may not introduce others unexpected side effects.
From there we discussed two options (but there may be others):
• Review and merge the PR [4], then later run some intensive tests aiming to verify
[2],[3] and checking the robustness in order to reenable NS
I think this is the best solution. If NS is “not working” today, then merging the PR won’t
make it “not work” any less ;)
Given we have a reliable way to disabled it at build time, I think that merging, testing,
and then eventually discussion and decision of enabling by default is the best plan.
• Build some tests for
• measure the benefit of NS as [2] and [3] do not prevent some performance tests
• identify possible reproducers for [2] and [3]
• create robustness and long duration NS specific tests
• review and merge the PR [4]
As PR [4] is not intended for perf improvement, the step 2.1 will impact the priority
according to the performance benefits.
I think that the test plan is good. 2 and 3 will be hard to reproduce I think, they seem
like they require complex state and interactions. The goal of the PR I have open is to
reduce the complexity of NS integration into DS, so that tests for situations like 2 and 3
can be more easily created (and hopefully the simpler integration is going to resolve some
issues).
The “benefit” of NS is hard to say - because the other parts of the server core are still
needing threading improvement, NS may help some aspects (connection acceptance performance
for example), we won’t see all the of the benefits until we improve other parts. There are
many bottlenecks to fix! An example is lib globs high use of atomics (should be COW), and
the plugin architecture.
At a theoretical level, NS will be faster as we don’t need to lock and poll over the
connection table, but today the CT may not be the bottleneck, so NS may not make this
“faster” just by enabling.
Comments are welcomed
Regarding 2.1 plan we made the following notes for the test plan:
The benefit of Nunc-Stans can only be measure with a large number of connections (i.e.
client) above 1000. That means a set of clients (sometime all) should keep their
connection opened. Clients should run on several hosts so that clients are not the
bootleneck.
For the two types of events (new connection and new operations), the measurement could
be
• Event: New connections
• Start all clients in parallel to establish connections (keeping them opened) take the
duration to get 1000, 2000, ... 10000 connections and check there are drop or not
• Establish 1000 connections and monitor during to open 100 more, the same starting
with 2000, 10000
• Client should not run any operations during the monitoring
• Event: New operations
• Start all clients and when 1000 connections are established, launch simple operations
(e.g. search -s base -b "" objectclass) and monitor how many of them can be
handled. The same with 2000, ... 10000.
• response time and workqueue length could be monitored to be sure the bottleneck are
not the worker.
At one point I started (but did not finish) ldclt integration with lib389. It would be
great to have a series of stress and acceptance tests like these available in lib389 for
our release validations, so I believe that their creation is valuable regardless of NS.
—
Sincerely,
William