I have a customer with a strange problem - they get very poor performance
when reading from an NSS volume on an external storage cabinet. But
writing gives good performance, as does reading from volumes on the
internal disks. And this does not just happen on a single server - they
have four that have the same problem. The thing is, it works fine on
Windows 2003...
Anyway, the setup is like this:
Server: Dell Poweredge 1850 with 4GB RAM
Ctrlr: PERC 4e/SI for internal disks (256MB cache)
PERC 4e/DC for external disks (128MB cache, FW 521X)
OS: NetWare 6.5 with SP 4a / NetWare 6.5 with SP 5
Volumes: SYS: (NSS) on PERC 4e/SI
DATA05: (NSS) on PERC 4e/SI
VOL01: (NSS) on PERC 4e/DC
- According to the customer, Dell says no more performance is gained if
you go above 8 spindles. So the Perc 4e/DC has 14 drives attached, which
are split into two raid arrays of 7 and 6 disks with one global spare.
Each raid array then has one logical drive, containing one NSS partition.
And VOL01: is on a pool consisting of both of these partitions.
- The customer installed the OS himself, and has told me that the only
two parameters he has changed besides timesync and SLP are:

Using TSATEST I got about 4.5GB/minute when copying from VOL01: (i.e. the
volume attached to the external storage cabinet).
Using Total Commander and copying from the NetWare server to a Windows
2003 server we get performance as follows:
- Copying from the SYS: volume on NetWare to the W2K3 server: 70MB/s
- Copying from the VOL01: volume on NetWare to the W2K3 server: 10-22MB/s
- Copying from the E: drive on the W2K3 server to the VOL01 volume: 80-
Using Compaq File Manager I get about 4GB/minute (i.e. about 65MB/s) when
copying between a volume attached to the internal RAID controller and one
attached to the external controller (both ways).

Things I have tried:
- I increased the MinCacheBuffers from 512 to 2048, which gave a small
performance increase
- I increased the ReadAheadBlks parameter from 2 to 20, which gave a
small performance increase
- I enabled local file caching (SET Client File Caching Enabled = On),
which gave a huge performance increase (up to 9.5GB/minute with TSATest,
for instance) but this gave Total Commander a "kangaroo copy" (i.e. it
copied, then stopped for a long while before starting again)

An explanation for the excellent write performance would be that the PERC
4e/DC controller was geared towards writing rather than reading, but this
does not explain how TSATest and CPQFM gives both good read and write
Another explanation would be that the network driver could be at fault,
but I have tried upgrading the driver (CE1000.LAN v. 7.65 to E1000.LAN v.
8.14) which did nothing. But I will ask the customer to try another NIC,
when he has a chance.

I intend to do some measurements using IOMeter today, which I will post,
but would appreciate any help that can be had in this.

Ragnar Storstrm
Stover AS