I am looking at setting up a system to support analysis of large
datasets. The application normally reads through the source data,
writes out a working temporary, and then continues to make read passes
through the work file. Occasionally it will read from the work files
and write out a new work file. The files will be >100GB. The I/O
should mainly be sequential with read passes being far more common.

The hardware we have to work with is a Dell 6850 server with a SAS 5/E
adapter (4x3.0 Gbps port) and a MD3000 RAID w dual 4x3.0 Gbps port
(the dual allows redundancy but we don't need that now). The MD3000
is loaded with 15x300GB 15K drives. The MD3000 is rated "up-to"
1400MB/sec. The Dell is loaded with two 7140M quad-core processors w/
16MB of cache each and 64GB of RAM. The system will be running
Windows 2003 x64. For better or for worse, this is fixed in the short-
term. We can buy different controller cards or minor software (<
$2000) if we could expect marked advantages.

My first question is about multipath/load balancing -- will I need a
multipath/load balancing solution (e.g. Dell/EMC PowerPath) in order
to get maximum I/O performance with this configuration? I am not
familiar with these multilane SAS cards -- will each path appear as a
separate HBA to the host or will the SAS 5/E effectively trunk the
four lanes into one 12Gbps channel? When I worked with multiport FC
cards in the past, each FC port operated as a separate controller and
the OS/drivers had do load balancing across the ports in order to
achieve the highest performance.

As for data layout, I am considering breaking the RAID in to three
LUNs: 5xdisks in RAID5 for primary data, 5xdisks in RAID5 for working
datasets, 4xdisks in RAID0 for sort scratch, and 1 for hot-spare
(which isn't strictly necessary since this will not be a 24x7
system). However, that layout will limit me to 400MB/sec reading.

Alternatively, I wonder if I wouldn't be better off with all 15 disks
in a single RAID5 LUN. Could I hit 1000MB/sec sequential read in this
configuration? Should I worry about spindle contention during
simultaneous sequential read/write or can the RAID controller cover
that up? I am more familiar with UNIX systems connected to dumber
hardware. What is the current performance envelope of NTFS on W2K3
with modern hardware?

Any thoughts or suggestions appreciated

P.S.The e-mail address in the From: address is not active. Please
reply to the newsgroup.