Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Coretime interface #5

Merged
merged 11 commits into from
Jul 25, 2023
176 changes: 176 additions & 0 deletions text/0005-coretime-interface.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,176 @@
# RFC-5: Coretime Interface

| | |
| --------------- | ------------------------------------------------------------------------------------------- |
| **Start Date** | 06 July 2023 |
| **Description** | Interface for manipulating the usage of cores on the Polkadot Ubiquitous Computer. |
| **Authors** | Gavin Wood |
gavofyork marked this conversation as resolved.
Show resolved Hide resolved


## Summary

In the Agile Coretime model of the Polkadot Ubiquitous Computer, as proposed in RFC-1 and RFC-3, it is necessary for the allocating parachain (envisioned to be one or more pallets on a specialised Brokerage System Chain) to communicate the core assignments to the Relay-chain, which is responsible for ensuring those assignments are properly enacted.

This is a proposal for the interface which will exist around the Relay-chain in order to communicate this information and instructions.

## Motivation

The overall motivation for splitting out functions from the Relay-chain onto System parachains is well understood. An well-understood interface is necessary for ensuring multiple chains are able to coordinate their efforts.
gavofyork marked this conversation as resolved.
Show resolved Hide resolved

## Requirements
gavofyork marked this conversation as resolved.
Show resolved Hide resolved

- The interface MUST allow the Relay-chain to be scheduled on a low-latency basis.
- Individual cores MUST be schedulable, both in full to a single task (a ParaId or the Instantaneous Coretime Pool) or to many unique tasks in differing ratios.
- Typical usage of the interface SHOULD NOT overload the VMP message system.
- Worst case usage of the interface MUST NOT cause the Polkadot system to fail.
- The interface MUST allow for the allocating chain to be notified of all accounting information relevant for making accurate rewards for contributing to the Instantaneous Coretime Pool.
- The interface MUST allow for Instantaneous Coretime Market Credits to be communicated.
- The interface MUST allow for the allocating chain to instruct changes to the number of cores which it is able to allocate.
- The interface MUST allow for the allocating chain to be notified of changes to the number of cores which are able to be allocated by the allocating chain.

## Stakeholders

Primary stakeholder sets are:

- Developers of the Relay-chain core-management logic.
- Developers of the Brokerage System Chain and its pallets.

_Socialization:_

This content of this RFC was discussed in the Polkdot Fellows channel.

## Explanation

The interface has two sections: The messages which the Relay-chain is able to receive from the allocating parachain (the *UMP message types*), and messages which the Relay-chain is able to send to the allocating parachain (the *DMP message types*). These messages are expected to be able to be implemented in a well-known pallet and called with the XCM `Transact` instruction.

Future work may include these messages being introduced into the XCM standard.

### UMP Message Types

#### `request_core_count`

Prototype:

```
fn request_core_count(
count: u16,
)
```

Requests the Relay-chain to alter the number of schedulable cores to `count`. Under normal operation, the Relay-chain SHOULD send a `notify_core_count(count)` message back.

#### `request_revenue_info_at`

Prototype:

```
fn request_revenue_at(
when: BlockNumber,
)
```

Requests that the Relay-chain send a `notify_revenue` message back at or soon after Relay-chain block number `when` whose `until` parameter is equal to `when`.

#### `credit_account`

Prototype:

```
fn credit_account(
who: AccountId,
amount: Balance,
)
```

Instructs the Relay-chain to add the `amount` of DOT to the Instantaneous Coretime Market Credit account of `who`.

It is expected that Instantaneous Coretime Market Credit on the Relay-chain is NOT transferrable and only redeemable when used to assign cores in the Instantaneous Coretime Pool.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How do I get my credits back that I don't want to use in the market anymore as credits?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

you can't


#### `assign_core`

Prototype:

```
type PartsOf57600 = u16;
enum CoreAssignment {
InstantaneousPool,
Task(ParaId),
}
fn assign_core(
core: CoreIndex,
begin: BlockNumber,
assignment: Vec<(CoreAssignment, PartsOf57600)>,
gavofyork marked this conversation as resolved.
Show resolved Hide resolved
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To be clear, the caller here doesn't have the ability to provide information about how these resources are allocated - as in, whether multiple candidates from multiple paras are included every relay chain block, or if they take turns on relay-chain blocks.

That doesn't seem to be required in #1 or #3 , but if that needs to be specified later on then we'd have to adjust this interface in a future RFC.

end_hint: Option<BlockNumber>,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Noting here - the main purpose of the None end-hint is that it allows the broker chain to skip sending this message sometimes.

Some users of #1 will likely split their regions up and defer their allocate calls until the last possible moment. When the next-up region has the same assignment as the previous, there's no need to send a new assign_core message

)
```

Requirements:

```
assert!(core < core_count);
assert!(targets.iter().map(|x| x.0).is_sorted());
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Targets here means assignment?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If these are the assignments, why do they need to be sorted?

Copy link
Contributor

@rphmeier rphmeier Jul 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

They don't strictly need to be (depends on implementation), but it's better to start with a more constrained interface

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I mean it will depend on the implementation of how to sort CoreAssignment, but this will probably mean that lower task numbers will be scheduled earlier?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I really don't get the need for sorting it. Especially when we want to express something like Task(1), Task(2) should both use 1/2 of the same core at the same block.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

i have no opinion at all here.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Given the comments from @rphmeier below, I see why the sorting doesn't has any effect

assert_eq!(targets.iter().map(|x| x.0).unique().count(), targets.len());
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But doesn't this prevent interleaved execution of two parachains? Because then we would have multiple times the same CoreAssignment::Task in this list?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Interleaving would be something like

targets: Vec<(Task(1), 1/2), (Task(2), 1/2)>

assert_eq!(targets.iter().map(|x| x.1).sum(), 57600);
```

Where:
- `core_count` is assumed to be the sole parameter in the last received `notify_core_count` message.

Instructs the Relay-chain to ensure that the core indexed as `core` is utilised for a number of assignments in specific ratios given by `assignment` starting as soon after `begin` as possible. Core assignments take the form of a `CoreAssignment` value which can either task the core to a `ParaId` value or indicate that the core should be used in the Instantaneous Pool. Each assignment comes with a ratio value, represented as the numerator of the fraction with a denominator of 57,600.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't get how assignment should work? How does the assignment and fraction map to scheduling the assignment to the core?

I mean for things like [(Task(1), 0.5), (Task(2), 0.5)] task 1 and task 2 will be running interleaved all the time?

Copy link
Contributor

@rphmeier rphmeier Jul 20, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How does the assignment and fraction map to scheduling the assignment to the core?

It means that the relay-chain should allow the task/assignment to use that proportion of the core's resources. It's better to leave it general like this, because this definition also allows tasks to use 1/2 of the core every relay-chain block, when we get support for having multiple candidates per core at a time.

And scheduling tasks to specific relay-chain blocks is actually really bad - this formulation does work nicely with e.g. #3 , which is a probabilistic scheduler. Giving access to core resources in expectation is the best way to do it, for both user experience and system utilization.

I mean for things like [(Task(1), 0.5), (Task(2), 0.5)] task 1 and task 2 will be running interleaved all the time?

Yes, or (Task(1), 0.25), (Task(2), 0.5), (Instantaneous, 0.25) would mean that:

  • Task 1 gets to use 1/4 of the core's resources for this timeframe (1 block out of every 4 relay chain blocks)
  • Task 2 gets to use 1/2 of the core's resources (2 blocks out of every 4)
  • Instantaneous gets to allocate 1 block every 4 relay chain blocks.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's better to leave it general like this, because this definition also allows tasks to use 1/2 of the core every relay-chain block, when we get support for having multiple candidates per core at a time.

But how would that look like? Currently the fraction expresses how much blocks a task can build for a given time range. I assume this time range will be TIMESLICE and both sides relay chain/brocker chain will be aware of this?

We will need something to tell the relay chain that a certain task actually needs only 1/2 of a core and not 1/2 of the time range.

Copy link
Contributor Author

@gavofyork gavofyork Jul 21, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's designed to be probabilistic - over a sufficiently long period it would be scheduled 1/2 of the time, and the Relay-chain is expected to be as fair as it can be, but practical constraints preclude the ability to state anything that is "perfectly fair all the time". We don't explicitly specify a timerange - that could push the relay-chain to be instructed to do things it's not practically capable of doing; the relay-chain is expected to just maximise the fairness over a minimal time range.

Practically speaking over the next 24 months of our technology, the timerange over which we'd expect the fairness to play out would be 80 relay-chain blocks. This timerange could possibly reduce as scheduling becomes tighter.

Copy link
Contributor

@rphmeier rphmeier Jul 21, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But how would that look like? Currently the fraction expresses how much blocks a task can build for a given time range. I assume this time range will be TIMESLICE and both sides relay chain/brocker chain will be aware of this?

While it would be valid for the relay-chain side to interpret this interface as giving each chain 1/2 of the timeslice (e.g. A gets an hour of uninterrupted time, then B gets an hour), that would be a pretty bad scheduler. Scheduler implementations (like #3) should probably aim for probabilistic interleaving while minimizing starvation.


If `end_hint` is `Some` and the inner is greater than the current block number, then the Relay-chain should optimize in the expectation of receiving a new `assign_core(core, ...)` message at or prior to the block number of the inner value. Specific functionality should remain unchanged regardless of the `end_hint` value.

On the choice of denominator: 57,600 is a very composite number which factors into: 2 ** 8, 3 ** 2, 5 ** 2. By using it as the denominator we allow for various useful fractions to be perfectly represented including thirds, quarters, fifths, tenths, 80ths, percent and 256ths.

### DMP Message Types

#### `notify_core_count`

Prototype:

```
fn notify_core_count(
count: u16,
)
```

Indicate that from this block onwards, the range of acceptable values of the `core` parameter of `assign_core` message is `[0, count)`. `assign_core` will be a no-op if provided with a value for `core` outside of this range.

#### `notify_revenue_info`

Prototype:

```
fn notify_revenue(
gavofyork marked this conversation as resolved.
Show resolved Hide resolved
until: BlockNumber,
revenue: Balance,
)
```

Provide the amount of revenue accumulated from Instantaneous Coretime Sales from Relay-chain block number `last_until` to `until`, not including `until` itself. `last_until` is defined as being the `until` argument of the last `notify_revenue` message sent, or zero for the first call.

This explicitly disregards the possibility of multiple parachains requesting and being notified of revenue information. The Relay-chain must be configured to ensure that only a single revenue information destination exists.

## Performance, Ergonomics and Compatibility

No specific considerations.

## Testing, Security and Privacy

Standard Polkadot testing and security auditing applies.

The proposal introduces no new privacy concerns.

## Future Directions and Related Material

RFC-1 proposes a means of determining allocation of Coretime using this interface.

RFC-3 proposes a means of implementing the high-level allocations within the Relay-chain.

## Drawbacks, Alternatives and Unknowns

None at present.

## Prior Art and References

None.