NIS?

Donald Becker becker at scyld.com
Wed Oct 3 06:47:42 PDT 2001


On Tue, 2 Oct 2001, Bob Campbell wrote:

> NIS or LDAP or normal passwd files or something else?

As a general rule, cluster compute nodes should avoid any network
directory/name service.

> does the ease of maintain the NIS passwd's make it worth
> the performance hit? how large is the performance hit?

The performance hit of a network directory service is minor with
workstation-like use, but when used with a cluster you have just created
a significant serialization point.  The server will be heavily loaded
with requests with every job initialization, usually just when it's busy
with other administrative work.

As a specific example, an NIS server will overflow its request queue
when a few hundred synchronized requests arrive.  The dropped queries
will timeout and retry.  This would have a minor impact on just a few
machines with decoupled use, but most cluster applications start out by
waiting for all compute node to finish initization.


The typical Linux C library selects the directory/name service based on
the configuration in /etc/nsswitch.conf.

Scyld solves the serialization problem by implementing a new
directory/name lookup services, and selecting our specific system in
nsswitch.  We use almost-numeric names e.g. ".1" for hostnames, and send
user/group information along with each job.

The best-practice approach for other cluster types is to specify "files"
(use local files) in nsswitch, and to use rdist/rsync to periodically
update the /etc/{passwd,group,hosts} files.

Donald Becker				becker at scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993





More information about the Beowulf mailing list