Hardware Reference
In-Depth Information
FIGURE 5.3 (See color insert): Block diagram of the ZFS-based Lustre file
system for LLNL's Sequoia system. The 96,000 compute node system with
768 I/O nodes is served by a 55-PB ZFS-based Lustre file system composed
of 768 OSTs and a metadata server with single failover.
given compute node is always serviced by the same I/O node (by default,
i.e., in the absence of a hardware/software failure which can be covered by a
failover). Also, a 128-compute-node partition is serviced exclusively by an I/O
node (by default, again); and the I/O node is running a full Linux operating
system, based on Red Hat Enterprise Linux. The Lustre client software is
a part of the I/O node operating system; and the I/O node talks to the file
system through an InfiniBand network, ZFS-based Lustre, TLCC2 OSS nodes,
and NetApp RAID controllers and disks.
5.6 Experience with ZFS-Based Lustre R and Sequoia in
Production
To call the file system component of the Sequoia BG/Q installation in LC
a major effort would be a profound understatement. The need for a new back-
end file system was anticipated well in advance of the Sequoia installation
date and development efforts were committed early on. Multiple transi-
 
Search WWH ::




Custom Search