提高unixware文件系统性能

1. Add the entry 'delaylog=nolog' to the mount options in /etc/vfstab, for example:

/dev/filesys /dev/rfilesys / vxfs 1 no delaylog=nolog,mincache=closesync SYS_RANGE_MAX

This will run your filesystem with no transaction logging, making it difficult to repair in the event of a system crash. However, it will improve performance, and if the files in this filesystem are considered to be temporary, then this option can be used.

More information on these options are available from the command:

man mount_vxfs

2. The above online manual page refers to the various 'mincache' options. The default is 'mincache=closesync' which is the slowest, but more reliable method.

However, adding the entry 'mincache=tmpcache' to the mount options in /etc/vfstab will cause your filesystem to become unmountable unless you are using the Advanced VxFS filesystem, On-Line Data Manager (ODM).

NOTES: In order to achIEve the best performance, it's helpful to look at the type of disk configurations available:

No Raid - Disks independent, no resilience

Raid 0 - Fast as disks are striped. 1 disk min, max depends on controller. No Fault Tolerance. Increase disk capacity.

Raid 1 - Mirrored disks. 2 disks required. Fault Tolerance. Hot Spare available.

Raid 5 - Resilience but slow as parity is spread across all the disks. 3 disk min, max depends on controller. Fault Tolerance. Hot spare available.

Raid 10 - An entire Raid 5 configuration mirrored. 6 disks min, max depends on controller. Fault Tolerant. Hot Spare available. Very slow but Excellent resilience. Rarely used.

Raid 0+1 - Raid 0 configuration but each disk is mirrored. 2 disks min, max depends on controller. Hot Spare available. Costly, because if you wanted a 100GB Logical Drive made from 10GB disks you would need 10 disks for the 100GB, and then a further 10 disks for the mirror.

Bus Speed - There are three components: the controller card, PCI bus inside the server, and the driver for the controller. All three need to be 64 bit to get true 64 bit, otherwise there will be a 32 bit bottleneck.

Controller - If the controller has memory with a battery backup then 'Write Back' can be used to use the full benefit of the on-board cache of the controller. The other option is the 'Write Thru' option which is slower but more reliable.

NOTE: If the database used is transaction based, then data can be written to the disk from the cache during a boot up if the power had gone down earlier. If the database is not fully opened then the transaction logs can get confused.

You may want to have separate controllers for separate filesystem activities.

You can also gain performance by ensuring that if the controller is multi-channel, each logical drive uses physical disks on the same channel.

It is recommended to have a separate controller for tape(s) and CDROM(s), rather than use the same controller as the disks. It is also recommended to not have too many SCSI devices on the same bus.

Sizes - You can set the Stripe Size on the controller and the Block Size of the filesystem to be the same to optimize performance for large file processing with large databases, such as Oracle, especially if these databases are on a separate controller. This would mean increasing the blocksize of the filesystem, when it is created, to the same size as the Stripe Size.

The larger the block size is, then each file created will consume that amount in the inode table. For example, a block size of 8k will consume 8k in the inode table for each file created. A typical example of this would be an Oracle database with a 8k block size setting within Oracle's configuration. You would make Oracle and the filesystem's block size consistent.

For some servers you may want to lower the Stripe Size if lots of small reads and writes are taking place so the controller is not waiting for its cache to be filled each time. However, the more read/writes there are, the more times an interrupt will be created for the action.

The norm though is to use the default RAID configuration setting for the RAID Stripe Size and the default filesystem block size for UnixWare 7, which is 1024.

Disks - In general, the more disks, the better performance as more spindles/heads are being used. The faster the disk revolution, the better. You should also check the cache provided by the disk from the disk's manufacturer.

Filesystems - Assign individual filesystems to separate tasks to reduce activity to the inode tables. In general, accept the number of inodes recommended by UnixWare 7 when adding the filesystem, even when you are expecting there to be a large number of small files written there. In this case, you would normally expect to double the number of recommended inodes.

However, the 'diskadd' command will actually give you an inode value of unlimited.

To check these settings, run this command on the filesystem (slice) in question:

mkfs -m /dev/rdsk/cXbXtXdXsX

Please note: During the Initial System Load (ISL), UnixWare2 & UnixWare 7.0.x will provide an Advanced option during the creation of the filesystems, in the Customise Filesystems option. By default, these 'vxfs' filesystems will have a 64K inode limit but this can be changed to 'Unlimited' in this Advanced option.

In the case of UnixWare 7, the best compromise for resilience and performance is to use Raid 1 for the root filesystem on one Logical Disk.

For each filesystem required, have a Logical Drive consisting of the necessary Raid 0+1 disks required for the size. If the number of disks is a factor, which is the norm, then have a reduced number of Logical Drives splitting the filesystems between them.

For any server, this value may need to be altered to find the best optimization, if the server is doing a combination of database and i/o work.

General speed testing, other than installing and testing your application, would be:

dd if=/dev/zero of=/<filesystem>/tmp/testfile bs=1024 count=5000000

(creates a 5GB file)

This will monitor the timings of any differences set; use 'sar -d' to monitor disk activity of each logical disk.

Also, before the server goes into production, remove disk(s) from the RAID configurations to ensure that the RAID controller detects that a disk has failed and that when it is re-inserted the RAID is successfully rebuilt.

Once in production, you may wish to look at the disk fragmentation options available with the 'fsadm' command, eg.:

fsadm -p 1 -e -s -v /<filesystem> fsadm -p 1 -D -d -s -v /<filesystem>

NOTE: Even if you optimally tune your vxfs filesystem, you might still encounter performance problem when restoring from tape a large number of small files. This overhead is only due to the vxfs design. For instance, while using cpio to restore around 300,000 files with size of 4-14k, it has been reported that a drop in performance was experienced from a 1-2M transfer rate (tape specification) to a 500K transfer rate.