AFS Server Performance Comparisons

1
AFS Server Performance Comparisons

• Bo Tretta
• Kim Kimball
• Jet Propulsion LaboratoryInformation Services -
  FIL Service
• http//fil.jpl.nasa.gov
• SLAC AFS Best Practices Workshop
• March 24, 2004

2

• Performance benchmarks of various hardware
  configurations
• Range of ages for hardware
• Different AFS versions
• Most hardware is already in operation and limits
  the testing we can perform
• Partitioned network
• Testing was performed on both sides of the
  network firewall.
• Operational Implications

3
Can we use less expensive hardware and still meet
performance goals?
4
Cell Configuration

• At the Jet Propulsion Laboratory in Pasadena, the
  AFS cell is configured with a firewall that
  splits the database servers as well as the
  fileservers.

5
Internet
Clients
Test Cell
Firewall
Flight Ops Firewall
afs06, afs07, afs20
Database Servers
File Servers
File Servers
Clients
Clients
File Servers
Database Servers
Database Servers
6
Benchmarking

• First assessment of OpenAFS fileserver hardware
  using Andrew Benchmark.
• Initial goal Determine if further assessment of
  inexpensive fileservers is warranted without
  wasting time and resources in the initial trials.

7
Methodology

• Compare Andrew Benchmark results from inexpensive
  Intel-based fileservers with results from
  existing Sun Solaris fileservers.

8
(No Transcript)
9
Performance test from a client outside of the
firewall
10
Observations

• afslinux02 was not performing the same as
  afslinux01 which is identical hardware.
• After examining afslinux02, it was found that the
  L2 cache module was not installed.
• The subsequent tests were performed with the L2
  cache module installed.

11
Performance test from a client outside of the
firewall with L2 cache installed
12
Performance test from a client inside of the
firewall. Did not test to the test cell systems
because the production servers can not be
modified.
13
Conclusions

• Inexpensive hardware for OpenAFS fileservers is
  not ruled out.
• Follow on Proceed to stress testing to determine
  feasible transaction rates.

14
Examining a Myth

• Expensive big iron is frequently purchased
  because we know it will get the job done.
• The result can be a collection of a small number
  of expensive fileservers.

15
But

• A larger number of inexpensive fileservers may
  provide equivalent performance.
• It may be that the initial investment in the
  larger number of less expensive fileservers does
  not significantly differ from the smaller
  collection of big iron for a given aggregate
  transaction load.

16
Benefits

• Initial cost aside, the more and cheaper
  fileserver approach offers
• Inexpensive incremental increase in capacity.
• Cost effective redundancy.
• Better manageability.

17
Cost effective redundancy

• If a fileserver fails, it can be immediately
  replaced, with similar (or identical) hardware
  kept for this purpose.

18
Better manageability

• We theorize that it will be easier to take a
  machine out of service when it houses less RW
  data.

19
Assumption

• That the time required to move the Read/Write
  volumes in the OpenAFS namei implementation does
  not increase to a point that volume moves are
  truly abysmal.

20
Benefit

• It will be easier to recover from unexpected
  hardware failure. Its easier to justify a
  spare server at 10,000 than at 100,000.

21

• Contact Information
• Bo Tretta botretta_at_jpl.nasa.gov
• Kim Kimball Kim.Kimball_at_jpl.nasa.gov