Hub/Switch Installation Guide

Chapter 3 System Preparation
HPSS Installation Guide September 2002 159
Release 4.5, Revision 2
To test whether an IP address is reachable (non-zero exit status indicates the ping was not
successful):
% ping -c 1 <ipAddress>
• Determine which networks will be used for control vs. data paths. DCEshould notuse all
available networks on amulti-homed systemunless each of those networks is guaranteed
to haveconnectivity toother DCE services. If aparticular networkis removed (physically,
or routing is changed), that connection remains in DCE's RPC mappings. DCE will
continue to try and use that network despite the loss of connectivity. The timeouts and
retries can have an adverse affect on HPSS as well as other applications relying on DCE.
◆ Isolate DCE communication to the designated control network(s) on all HPSS, DCE,
and SFS nodes in order to separate the HPSS control and data paths.
◆ As desired,tell DCE to ignorecertain networkinterfaces in orderto separatethe HPSS
control and data paths. For example, to ignore FDDI (fd0) and HIPPI (hp0), add the
following line to /etc/environment:
RPC_UNSUPPORTED_NETIFS=fd0:hp0
• Place allHPSS,SFS,andDCE servermachine IPaddresses inalocalhosttable (/etc/hosts).
ForAIX,configurethemachinetouse thetableas backupintheeventofaDNSfailure.The
file /etc/netsvc.conf should be modified to look like the following:
hosts=bind,local
The netsvc.conf file is used to specify the ordering of host name resolution. In the above
ordering, DNS will be used first, if the host name is not found, then the local /etc/host file
will be used.
4. For each ethernet network interface, verify that both the en0 and et0 interfaces are not
configured at the same time (we recommend only using en0 unless the other machines in
the networkareall usingthe 802.3et* interface). Configurethe localname servicewith the
unique hostnamefor each networkinterface on all nodesand verify thateach hostname is
resolvable from other nodes.
To understand and optimize the operation of HPSS, some baseline measurement, evaluation, and tuning of
the underlying network and IO subsystems is necessary. Appropriate tuning of these components is
necessaryforHPSStoperform asexpected.Anyonecomponent thatisnotperformingat itsoptimalratewill
have an adverse affect on the performance of the entire system. The steps and tools listed here will help a site
make the best use of their available resources. The measurements taken should be saved for future reference.
If performance problems occur later on, these values can be compared with new measurements to determine
if the performance problems are related to changes in the subsystems. Though disk and tape configurations
rarelychange withoutan administrator'sknowledge, networkscan "mysteriously"down gradefor avariety
of reasons. This is especially true for client access to the HPSS system.
• Verify that network TCP throughput has been optimized and the performance of each
network is at expected levels in both directions (especially check HPSS data networks
betweenMoversandbetweenMoverandclientnodes).Usingttcporanothernetworktool,
measuretheperformanceofallthe networksthat willbeused incommunications between
DCE, HPSS, SFS, and client machines. If multiple paths between machines are to be used,
then all ofthem need to be measuredas well. The transferrates for networks differgreatly
depending upon the individual technology and network settings used. It is important to