Resource Management

1
Resource Management

• Foreign Drivers
• Reserves

2
Goals

• Identify issues devolving from driver support,
  and try to shape a solution
• Architectural support for residency and timing
  requirements to be specified and met
• Better understand issues in CPU scheduling

3
Current Issues

• Supporting foreign (i.e. Linux) drivers will
  create a two-level resource management problem.
• Impact of this is unclear
• What to do is unclear
• EROS does not export multiplexing controls for
  logical to physical resource binding
• Residency contracts
• Scheduling contracts
• Need for real-time checkpointing

4
Foreign Drivers

• UNIX drivers assume
• Per-process kernel stack
• Existence of kmalloc()
• EROS now assumes
• Interrupt-style kernel
• No kmalloc()

• EROS will adopt the slab allocator
• Page-based kmalloc()
• Ageing issues
• EROS already has in-kernel processes
• These can export start capabilities
• They retain a stack

5
Foreign Drivers Proposal

• Proposal
• Use an in-kernel process for each foreign driver.
• Wherever possible, use asynchronous I/O.
• Build a clean handoff between the EROS I/O
  subsystem and the foreign drivers.
• This has more or less been done in OSKit
• Issues
• Error handling
• Scheduling
• Interaction with CPU scheduler
• Need for reserved scheduler slots?

6
Residency Reserves

• Plan
• Pool of residency colors, each w/ fixed lower
  bound of object frames (page, node frames)
• Every process, segment has a reserve set (similar
  to meter) that identifies its color
• Once resident, pages/nodes compete against their
  own colors for residency
• Questions
• Fixed number of colors, or subdivisible?
• Do colors need to be prioritized?
• Useful to have opportunistic watermarks?
• Is the model push or pull (Roscoe)
• Is this enough? Kmalloc()?

7
Semi-Shared Residency

• Shared object, critical memory region, frame
  buffer, etc.
• Scoping rules follow keeper logic nearest
  enclosing reserve

Reserve Pool
8
CPU Reserves

• Currently using a (broken) capacity reserve
  scheduler.
• Roscoe suggests that there may be a better
  design.
• Problem storage for the reserve table.

9
Real-Time Checkpoint

• We now have a checkpoint design with sub-msec
  snapshot phase
• Uses lazy propagation of ckpt generation bit
• Use Liedtkes infinite pinning design as a model
  for understanding COW impact

10
Design Session

• Storage Reserves
• What do memory reserves need to say?
• Can an infinite pool of colors be supported? Is
  it needed?
• What happens when memory is oversubscribed?
• CPU Reserves
• Discussion of CPU scheduling
• Discussion of data vs. control transfer.
• Foreign Driver Issues
• Impact of multi-tier allocation?