wiki:CreditProposal

Version 1 (modified by davea, 13 years ago) (diff)

--

Proposal for a new credit system

This is a discussion of (and tentative proposal for) a new credit system for BOINC.

Requirements

  • The system should reflect how much "utility" volunteers provide to projects. There are many types of utility (see below). Projects should be able to define utility however they want.
  • The system should limit the amount of credits that projects can grant, so that there will be no incentive for average volunteers to select projects based on credit.

Types of utility

Possible types of utility:

  • Computation (FP, integer, or mixed) without tight deadlines (this is typical of most current projects).
  • Computation with a tight latency bound (minutes or hours).
  • Computation that needs a large amount of RAM.
  • Computation that needs a large amount of storage.
  • Storage (e.g. GB/day)
  • Storage, with network bandwidth and/or availability requirements
  • Network bandwidth (e.g. web crawling)
  • Network bandwidth at particular times of day (e.g. Internet performance study)
  • Deployment on a wide range of computer types (e.g. studies of computer usage)
  • Computation with human "steering"
  • Human activity (e.g. [email protected] projects)

Proposal

  • Computers are described by a set of parameters (FP and int benchmarks, #CPUs, cache sizes, memory bandwidth, available RAM, available disk, presence of particular GPUs, network bandwidth, available fraction, connected fraction, maybe others).
  • Each project P publishes a "credit function" C(H) specifying, for a host H with given parameters, how much credit per day would be granted if H is attached exclusively to P.
  • Normalization rule: the average of C(H) over all hosts participating in BOINC must be about 100.
  • Accounting rule: the credit granted by a project P cannot exceed the sum over hosts H of
    RS(H, P)*C(H)
    
    where RS(H, P) is the fractional resource share of H's attachment to P.

The normalization and accounting rules would be evaluated by cross-project statistics sites.

Examples

  • Computational projects would have to derive a credit function based on how fast various types of computers run their applications (and how much they value each application). We can supply automated tools for this.
  • Suppose a project's application needs at least 16 GB RAM. Its credit function would be zero for hosts with < 16 GB RAM. Its value for hosts with at least 16 GB would be limited by the normalization rule.
  • For a storage-only projects, C(H) would be proportional to available disk space (possibly with some additional consideration for network bandwidth etc.).
  • Using the published credit functions, it would be possible to develop a "credit maximizer" web site, where users can enter the parameters of their computer, and it tells them how much credit/day each project would give them.
  • Similarly, it would be possible to develop a web site that tells users how much additional credit/day they could get by adding more disk/RAM/network bandwidth etc. to an existing computer.

Questions and comments

  • BOINC doesn't currently have code to measure memory bandwidth or cache sizes; we'd need to develop this.
  • Credit can no longer be used as a basis for FLOPS estimates; we'll need something else for that purpose (or keep around the existing credit system and use it only to estimate FLOPS).
  • Suppose a project's application needs X GB of RAM, and there is only one host in the BOINC population with that much RAM. Then its credit per day for that host can be 100N, where N is the size of the population. In other words, that computer can get as much credit as all other computers combined. Is this desirable?
  • The accounting rule doesn't take into account non-competing projects. E.g. suppose project A uses only CPU and project B uses only disk. If hosts attach to both projects with equal resource shares, the projects will be limited to issuing 50 credits/day on the average.
  • Suppose a project uses a resource that no other project wants (e.g. QCN uses the accelerometer found in some laptops). Should the project be able to allocate the full 100 credits/day on average?