Client setup

Client protocol extension

Install a suitable kernel module

• via rpm

% rpm -q kernel-module-openafs-`uname -r`
kernel-module-openafs-2.6.18-128.1.10.el5-1.4.7-68.2.SL5.i686

• or manually

  1. From the source directory copy

     src/libafs/MODLOAD-`uname -r`-*/libafs.ko

     • to

     /lib/modules/`uname -r`/kernel/fs/openafs/libafs.ko

  2. depmod

Restart of the AFS service

Upon startup of the afs service, the following should show up:

# service afs restart
...
# dmesg | grep patch
libafs with   'rxosd support' 'vicep-access' patches

(vicep-access is not necessary for RxOSD operation per se, but is required for making proper use of a Lustre or GPFS backend.)

Because afs restart doesn't work often, is is better to reboot the client.

Proper afs commands

Install a suitable openafs package

# rpm -q openafs
openafs-1.4.10.osd.vpa.r691-77alma.sl5.x86_64
# fs protocol
Enabled protocols are  RXOSD (1 parallel streams on connections with rtt > 10 ms).
# osd help
osd: Commands are:
...

Memory cache

When using a cluster filesystem backend, memcache has proven to be the fastest alternative (at least for Lustre via Infiniband).

In /etc/sysconfig/afs:

OPTIONS="-memcache -stat 8192 -chunksize 20 -daemons 8 -volumes 64"

This selects memcache and boosts the chunk size to 1MB. To avoid a low overall number of cache chunks:

CACHESIZE="262144"

256 MB seems a reasonable size.

Accessing a cluster filesystem backend

In Zeuthen, the Lustre symlinks is handled by the SLZ_vicep_lustre_links RPM.

Generally (e.g. Lustre):

• Make sure the shared vice partition is accessible:

  # ln -s /vicepal /lustre/...

  • or resp.

  # mount -t lustre 141.34.218.7@tcp:/zn_test /vicept

  • (see "Using a Lustre storage backend" below)
• make the client detect the vice partition

  # service afs-vicep-access restart
  ...
  # dmesg | tail
  ...
  Visible OSD 2 lun 37 cell 33 == ifh.de

Fileserver setup

Adding an rxosd service

It is sensible to have fileserver and rxosd processes share the same machines. This way, access to files even in a single volume can be spread to all fileservers (given that clients don't access the same (set of) file(s) all the time).

• create a dedicated vice partition

  # mkdir /vicepc
  <edit /etc/fstab>
  # mount /vicepc
  # touch /vicepc/OnlyRxosd

  • The last command ensures that the fileserver process will not claim this partition.

• create the service instance (with admin privileges or -localauth)

  # bos create <server-machine> rxosd simple /usr/afs/bin/rxosd

  • There is no harm in doing this even before creating a vice partition. Please note the discussion of Lustre storage backends below.

• add the new OSD to the OSDDB

  # osd createosd -id <id> -name <osd-name> -ip <machine-ip> -lun <lun> 1m 64g

  id

  must not have existed in the OSDDB before, check with osd l -all

  osd-name

  must also be unique

  lun

  numeric representation of the vice partition, /vicepa is 0, /vicepc is 2 and /vicepal is 37

  size constraints

  1m and 64g are arbitrary.

  Note that the minimum size directly influences what files get stored to OSD. If this is done before creating vice partition and/or service instance, the OSD will appear as "down" in the osd l output. No harm done.

Using a Lustre storage backend

Basically, this is the same process as adding a regular rxosd service. The vice partition resides in a Lustre filesystem, however. The rxosd process must be launched with a special parameter:

# bos create <server-machine> rxosd simple "/usr/afs/bin/rxosd -lustrehack"

A larger maximum file size for your Lustre-OSD is not required, the fileserver will never get to see a size that surpasses any client's cache size.

Lustre mountpoint

E.g.

# mount -t lustre 141.34.218.7@tcp:/zn_test /vicept
# touch /vicept/OnlyRxosd

The rest is identical to the above.

Symlink

If Lustre is mounted anywhere else:

# ln -s /lustre/vpxl /vicepal
# touch /vicepal/OnlyRxosd
# touch /vicepal/AlwaysAttach

The last command enables the use of a regular directory as a vice partition. The rest is identical to the above.

Verifying the correct setup of the OSD

Assuming both service instance and vice partition were available at the time of updating the database, osd l should contain something along the lines of

  2 zyklop20-al    10694 gb   0.0 % up         64  64     zyklop20 37 (1mb-2tb)

The "up" signifies that the OSD is ready.

A "down" entry does not have to be a problem. It can (among other things) mean

• an unreachable host

• a missing rxosd process on the host machine

  bos status <osd-machine>

• a missing vice partition on the host machine
• a wrong LUN

  bos getlog <osd-machine> RxosdLog
  ...
  Thu Jun  4 13:20:00 2009 [0] FiveMinuteCheckLWP: statfs for /vicepam failed with 2

  • Here, the LUN had been one too large initially.

Database servers

Packages

If appropriate openafs-server packages are in use, there is nothing to do.

# rpm -q openafs-server
openafs-server-1.4.10.osd.vpa.r657-74alma.2.sl5.x86_64
# rpm -qf /usr/afs/bin/osdserver
openafs-server-1.4.10.osd.vpa.r657-74alma.2.sl5.x86_64

For a minimum intrusion on existing DB servers, use the dedicated openafs-osdserver package:

% rpm -q openafs-server openafs-osdserver
openafs-server-1.4.7-68.2.SL5.x86_64
openafs-osdserver-1.4.10.osd.vpa.r689-76alma.sl5.x86_64
% rpm -qf /usr/afs/bin/osdserver
openafs-osdserver-1.4.10.osd.vpa.r689-76alma.sl5.x86_64

It has no dependencies and will install regardless of the rest of your local AFS installation.

Launching the OSDDB service

With admin privileges:

bos create <db-machine> osddb simple /usr/afs/bin/osdserver

Once the osdservers are up and running, clients should be able to see this:

% osd l
 id name(loc)     ---total space---      flag  prior. own. server lun size range
  1 local_disk                                 wr  rd                 (0kb-1mb)

The local_disk entry is a default.

Enabling OSD for Volumes ...

 # vos setfields osdtestvol2 -osd 1

Update Client and Server

• At First:
  • check if all packages are aviable and that no errors will be send because of a downtime or a not available service
• Stop AFS process
• On Client: .

• % /etc/init.d/afs stop

• On Server: (maybe Client process also running on a server!)

•  % /etc/init.d/afs-server stop

• Update the packages
• First Erase Kernel Package:

•  % rpm -e kernel-module-openafs-2.6.18-164.11.1.el5-1.4.11.osd.vpa.r901-80alma.sl5.x86_64

• Update all the other packages, Kernel-module should be installed too. Add openafs-server on a server machine. .

  % yum update openafs-authlibs-devel openafs-authlibs openafs-client openafs openafs-krb5 openafs-debug openafs-compat openafs-devel

• Start AFS process
• On Client:

• % /etc/init.d/afs start

• On Server: .

  % /etc/init.d/afs-server start

• Start Vicep-Access on the Client

• % /etc/init.d/afs-vicep-access status
          Enabled protocols are  RXOSD (1 parallel streams on connections with rtt > 10 ms).
  % /etc/init.d/afs-vicep-access stop
  % /etc/init.d/afs-vicep-access start
  % /etc/init.d/afs-vicep-access status
          Enabled protocols are  VICEP-ACCESS (with Lustre hack) RXOSD (1 parallel streams on connections with rtt > 10 ms).

• Check Configuration
• On Client: .

  % dmesg
  ...
  Visible OSD 2 lun 37 cell 33 == ifh.de

• % KUSL.pl

• dont want to install something . . On Server: . .

  %  bos status localhost -localauth
  Instance fs, currently running normally.
      Auxiliary status is: file server running.
  Instance rxosd, currently running normally.

• Or other Instances.