Problems of Dynamic Service Deployment

1
Problems of Dynamic Service Deployment
2
Motivation

• Dynamic Service Deployment
• Install and setup a service on-the-fly
• Shutdown and cleanup a service also on-the-fly
• Dynamically deploy a Condor submit node
• Why is the problem hard?

3
Primary Problems

• Bootstrap
• User Privileges of Bootstrap (non-root)
• Remote Control
• Reliable Startup and Shutdown
• Executable Architecture
• Cleanup
• File dependencies (libs, etc...)
• No knowledge of system

4
Problem Focus

• Bootstrap
• User Privileges of Bootstrap (non-root)
• Remote Control
• Reliable Startup and Shutdown
• Executable Architecture
• Cleanup
• File dependencies (libs, etc...)
• No knowledge of system

5
(No Transcript)
6
Straw man
Pid as a Unique Handle to a Process
7
Slow Pid Reuse
8
Slow Pid Reuse
9
Improved Straw Man
Pid and Creation Time as a Unique Handle
10
Properties of Machine Time
11
Properties of Machine Time
Collision Range
12
Bulletproof Solution
Pid and Creation Time Handle with Collision
Avoidance
13
Midwife Implementation
14
Isn't SIGKILL a reliable shutdown mechanism?
15
Reliable Shutdown Pseudo Code
while( alive totalTime lt timeout)
attemptShutdown() alive isAlive(composite_id)
if( alive ) sleep(5)
totalTime 5
16
(No Transcript)
17
Conclusion

• Available in Condor v 6.7.18
• uniq_pid_midwife
• uniq_pid_undertaker
• Technical Report Coming Soon

18
(No Transcript)
19

• PPID 7338
• PID 7339
• PRECISION 100
• TIME_UNITS_IN_SECS 100.000000
• BDAY 579893001
• CONTROL_TIME 0
• CONFIRM 579893308
• CONTROL_TIME 0