Constraints for P|N Spin Locks

P|N locks ensure that a request is blocked at most once by another request with lower priority at the expense that there is no immediate bound on the number of blocking higher-priority requests. In the following, we denote the locking priority of requests for resource `q issued by jobs of a task Tx as πx,q,

and the set of remote tasks with respect to T_i as τ^R:

τ^R, {Tx | P (Tx) 6= P (Ti) } .

We apply response-time analysis [19] on a per-request basis to obtain an upper bound on the delay encountered when issuing a single request for a resource

`q with priority π. For a resource `q and task Ti, let Wq^P|N(Ti, π) denote the smallest positive value (if any) that satisfies the following recurrence:

W_q^P|N(T_i, π) = S(`<sub>q</sub>, π) + LP (`_q, π) + 1 where (3) S(`_q, π) = X

Tx∈τ^R∧πx,q≤π

njobs(T_x, W_q^P|N(T_i, π)) · N_x,q· L_x,q and

LP(`q, π) = max

Tx∈τ^R{Lx,q|πx,q > π}.

The recurrence can be solved via fixed-point iteration. Since a bound on the delay exceeding T_i’s deadline cannot be used as part of (effective) constraints, the fixed-point iteration can be aborted if no fixed-point with W_q^P|N(T_i, π) ≤ d_i is found. If a solution for W_q^P|N(T_i, π) satisfying the above recurrence can be found, then Wq^P|N(Ti, π) bounds the delay of a single request. We formalize this property with the following lemma.

Lemma 10. Lett0 be the time a jobJi of taskTi attempts to lock resource

`_qwith locking priorityπ, and let t₁be the time thatJ_isubsequently acquires

`q. With P|N locks, t1− t0 ≤ W_q^P|N(Ti, π).

Proof. Analogous to the response-time analysis of non-preemptive fixed-priority scheduling. The response-time of J_i’s request—that is, the maximum wait time Wq^P|N(Ti, π)—depends on the sum of the maximum length of one lower-priority request LP (`<sub>q</sub>, π) and all higher-priority requests of all remote tasks issued during an interval of length Wq<sup>P|N</sup>(Ti, π), that is, S(`q, π). Hence, throughout an interval of length Wq^P|N(T_i, π), by definition of Wq^P|N(T_i, π),

resource `q is unavailable for at most Wq^P|N(Ti, π) − 1 time units. Thus, Ji’s request is served after at most Wq^P|N(Ti, π) time units after it was issued. 

In the constraints we establish for the analysis of P|N locks, we exploit two simple monotonicity properties of Wq^P|N(Ti, π), which we next state explicitly for the sake of clarity. First, Wq^P|N(T_i, π) is monotonic with respect to scheduling priority. That is, the wait time of a request for `q issued by a local higher-priority task T_h with the same locking priority π is no longer than the wait time of Ti’s request. (In fact, the per-request wait-time bound is independent of scheduling priority since jobs spin non-preemptably.) Formally,

∀Th ∈ τ^lh : W_q^P|N(Ti, π) ≥ W_q^P|N(Th, π). (4)

The second monotonicity property that we exploit pertains to the locking priority π: in a P|N lock, requests issued with higher locking priority natu-rally do not incur more spin delay than requests issued with lower locking priority:

π⁰ < π → W_q^P|N(T_i, π) ≥ W_q^P|N(T_i, π⁰). (5)

The above monotonicity properties enables us to use wait-time bounds in constraints computed with the minimum locking priority of any requests issued by local higher and lower priority tasks, respectively. To simplify the notation, we define

π_q^minLP , max

Tx∈τ^ll{π_x,q|N_x,q > 0} and π^minHP_q , max

Tx∈(τ^lh∪{Ti}){π_x,q|N_x,q > 0}

to be the minimum locking priority of any lower-priority and higher-priority task, respectively, on Ti’s processor that accesses the global resource `q. These two definitions are needed because J<sub>i</sub> might be delayed transitively due to requests of local tasks with locking priorities lower than Ti’s own locking priority. To obtain valid (and simple) constraints, we make the following two simplifications: first, for a given resource `q, we assume that Ji and all higher-priority jobs that preempt J_i issue requests with locking priority π_q^minHP (the lowest locking priority that any such job uses), and second, we assume that all local lower-priority jobs request `_q with locking priority π_q^minLP (again, the lowest locking priority used by any local lower-priority job). Both of these are safe assumption due to the monotonicity property stated in Equation (5). However, we note that these simplifications are a potential source of pessimism that could be avoided with a significantly more complicated analysis setup, which we leave to future work.

Given Wq^P|N(T_i, π) (i.e., if it exists), we can constrain the the number of requests for `q that can contribute to Ti’s spin delay. First, we consider requests issued with higher or equal priority.

Constraint 10. In any schedule ofτ with P|N locks:

∀P_k, P_k6= P (Ti) : ∀`q∈ Q^g: ∀Tx∈ τ (P_k), πx,q ≤ π_q^minHP:

N_x,qⁱ

X

v=1

X_x,q,v^S ≤ njobs(Tx,W_q^P|N(Ti, π^minHP_q ))·Nx,q·ncs(Ti, q).

Proof. Let R denote a request for a resource `_q by T_i or a local higher-priority task. By the definition of π_q^minHP, R has at least the locking priority π_q^minHP and, by Lemma 10 and monotonicity properties of Wq^P|N stated in Equations (4) and (5), is hence delayed by at most Wq^P|N(Ti, π_q^minHP) time units (note that Wq^P|N(T_i, π_q^minHP) ≥ Wq^P|N(T_h, π_h,q) if T_h ∈ τ^lh and π^minHP ≥ π ). During an interval of length W^P|N(T, π^minHP), jobs of a

remote task Tx issue at most njobs(Tx, Wq^P|N(Ti, π^minHP_q )) · Nx,q requests for `q. The stated bound follows as at most ncs(Ti, q) requests for `q with a priority of at least π_q^minHP are issued by T_i or local higher-priority tasks. 

Requests with lower priority cause Ji to incur (transitive) spin delay at most once for each request by T_i or a task in τ^lh.

Constraint 11. In any schedule ofτ with P|N locks:

∀`_q∈ Q^g: X

Tx∈τ^R πx,q>π_q^minHP

N_x,qⁱ

X

v=1

X_x,q,v^S ≤ ncs(T_i, q).

Proof. Suppose not. Then at least one request for global resource `q issued by T<sub>i</sub> or a local higher-priority task is delayed more than once by a request for `q from a different processor issued with a lower priority. However, by definition P|N locks ensure that each request is blocked at most once by a lower-priority request for the same resource. Contradiction. 

Next, we consider arrival blocking. The number of lower-priority requests that cause arrival blocking is bounded by A_q.

Constraint 12. In any schedule ofτ with P|N locks:

∀`q∈ Q^g: X

Tx∈τ^R πx,q>π_q^minLP

N_x,qⁱ

X

v=1

X_x,q,v^A ≤ Aq.

Proof. Suppose not. In case A_q = 0, by definition of A_q, T_i incurs transitive arrival blocking due to a request for `q, although no access for `q from a local lower-priority task causes T_i to incur arrival blocking, which is impossible.

In case Aq = 1, a request for `q with priority at least π^minLP_q is delayed

more than once by requests for `<sub>q</sub> issued on other processors with a locking priority of less than π<sup>minLP</sup><sub>q</sub> . However, with P|N locks, a request for a re-source `_q cannot be delayed by more than one lower-priority request for `_q.

Contradiction. 

Next, we constrain the arrival blocking due to requests with higher priority issued from other processors.

Constraint 13. In any schedule ofτ with P|N locks:

∀`q ∈ Q^g : ∀Tx ∈ τ^R, πx,q ≤ π^minLP_q :

N_x,qⁱ

X

v=1

X_x,q,v^A ≤ njobs(Tx, W_q^P|N(Ti, π_q^minLP)) · Nx,q· Aq.

Proof. Let R denote the request by a local lower-priority job that causes T_i to incur arrival blocking. By definition of π_q^minLP, R has a priority of at least π_q^minLP, and, by Lemma 10, is hence delayed by at most W_q^P|N(T_i, π_q^minLP) time units (note that Wq^P|N(Ti, π_q^minLP) ≥ Wq^P|N(T_l, π_l,q) if T_l ∈ τ^ll and π_q^minLP ≥ π_l,q). During an interval of length Wq^P|N(T_i, π^minLP_q ), jobs of a remote task Tx issue at most njobs(Tx, Wq^P|N(Ti, π_q^minLP)) · Nx,q requests for

`<sub>q</sub>. The bound follows as T<sub>i</sub> is arrival-blocked via `_q only if A_q = 1. 

This concludes our analysis of P|N locks. Together with the generic Con-straints 1–7, the P|N-specific ConCon-straints 10–13 define a MILP that bounds the maximum blocking incurred by any J_i. In the unlikely case that the re-currence given in Equation (3) does not converge for some `q, Constraints 10 and 13 that depend on the wait-time bound W<sub>q</sub><sup>P|N</sup>(T<sub>i</sub>, π) must be omitted from the MILP for this resource `q.

Next, we present the constraints for the analysis of PF|N locks.