Speed Up Your Replica Ring
Novell Cool Solutions: Tip
Digg This -
Posted: 22 Apr 1999
At a local university, the network administrator was stymied. He had upgraded a NetWare server with a new motherboard, including a fast Pentium processor, plenty of RAM, and a new wide SCSI controller. This server held the master replica of two partitions.
The administrator completed the hardware upgrade, booted the server, and started the partition joining operation at 5:00 in the morning to avoid slowing things down for the students, who started coming into the labs at 6:00. To his surprise, the partition joining operation took FFFooorrrEEEvvveeerrr. The calls started coming in to the help desk promptly at 6:00 from students, complaining that logging in took ten minutes or more. By 10:00, lab managers and department heads were complaining. When the partitions finally joined around noon, login times went back to normal but by that time, the administrator was sweating bullets and actually concerned about his job security.
What went wrong? The answer lies in the concept of replica rings. Consider the simplified diagram below showing the replica-holding servers at the university:
As you can see, the administrator had three replicas of each partition, which is Novells recommendation. But Partition A had a replica on a slow server (an old 486/33). Since a partition-joining operation passes massive amounts of data among all servers in the replica ring, the slow server served as a bottleneck.
What should the administrator have done? First of all, youve got to wonder why the slow server held a partition in the first place. Maybe a bindery application required a replica on the slow server. (Whenever a bindery context is specified for a server, that server must hold a replica of the partition that includes that context.) Maybe the replica was created automatically when the server was first installed.
Here are some recommendations:
- Move the bindery application, if there is one, to a faster server and delete the replica from the slow server. This should have been done before the partitioning operation.
- Reduce the number of replicas per partition to two before the partitioning operation.
- If the budget allows, upgrade the slow server.
- After the two partitions are joined, reduce the total number of replicas to three. (This may be impossible if bindery applications are running on more than three servers.)
Novell Cool Solutions (corporate web communities) are produced by WebWise Solutions. www.webwiseone.com