87 lines
5.1 KiB
Markdown
87 lines
5.1 KiB
Markdown
Can we take the most basic protocol that satisfies the most basic liveness property (obstruction freedom) and "upgrade" it to bounded wait freedom?
|
||
|
||
**Contention manager:** is an object that allows progress of processes by providing contention-free periods for completing their invocations. It provides 2 operations:
|
||
- `need_help(i)`: invoked by $p_i$ when it discovers that there is contention
|
||
- `stop_help(i)`: invoked by $p_{i}$ when it terminates its current invocation
|
||
|
||
**Enriched implementation:** when a process realizes that there is contention, it invokes need_help; when it completes its current operation, it invokes stop_help.
|
||
|
||
**Why is it different from lock/unlock?** Because this allows failures, and they can also happen in the contention-free period.
|
||
|
||
**PROBLEM:** to distinguish a failure from a long delay, we need objects called ***failure detectors***, that provide processes information on the failed processes of the system. According to the type/quality of the info, several F.D.s can be defined.
|
||
|
||
**Eventually restricted leadership:** given a non-empty set of process IDs X, the failure detector $\Omega_{X}$ provides each process a local variable `ev_leader(X)` such that:
|
||
1. *(Validity)* `ev_leader(x)` always contains a process ID
|
||
2. *(Eventual leadership)* eventually, all `ev_leader(X)` of all non-crashed processes of X for ever contain the same process ID, that is one of them
|
||
|
||
REMARK: the moment in which all variables contain the same leader is unknown
|
||
|
||
#### From obstruction-freedom to non-blocking
|
||
|
||
```
|
||
NEED_HELP[1..n] : SWMR atomic R/W boolean registers init at false
|
||
|
||
need_help(i) :=
|
||
NEED_HELP[i] <- true
|
||
repeat
|
||
X <- {j : NEED_HELP[j]}
|
||
until ev_leader(X) = i
|
||
|
||
stop_help(i) :=
|
||
NEED_HELP[i] <- false
|
||
```
|
||
|
||
**Theorem:** the contention manager just seen transforms an obstr.-free implementation into a non-blocking enriched implementation.
|
||
|
||
*Proof:*
|
||
By contr., $\exists \tau$ s.t. $\exists$ many (> 0) op.'s invoked concurrently that never terminate.
|
||
Let Q be the set of proc.'s that performed these invocations.
|
||
- by enrichment, eventually `NEED_HELP[i]=T` ($\forall i\in Q$) forever
|
||
- since crashed are fail-stop, eventually `NEED_HELP[j]` is no longer modified ($\forall j \not \in Q$)
|
||
- $\exists \tau' \geq \tau$ where all proc.'s in Q compute the same X
|
||
|
||
**Observation:** $Q \subseteq X$ (it is possible that $p_j$ sets `NEED_HELP[j]` and then fails)
|
||
|
||
By definition of $\Omega_{X}, \exists \tau'' \geq t'$ s.t. all proc.'s in Q have the same `ev_leader(X)`
|
||
- the leader belongs to Q, since it cannot be failed
|
||
- this is the only process allowed to proceed
|
||
- because run in isolation, it eventually terminates (because of obstruction freedom)
|
||
|
||
#### On implementing $\Omega$
|
||
It can be proved that there exists no wait-free implementation of $\Omega$ in an asynchronous system with atomic R/W registers and any number of crashes
|
||
- crashes are indistinguishable from long delays
|
||
- need of timing constraints
|
||
|
||
1. $\exists$ time $\tau_{1}$, time interval $\nabla$ and correct process $p_{L}$ s.t. after $\tau_{1}$ every two consecutive writes to a specific SWMR atomic R/W by $p_{L}$ are at most $\nabla$ time units apart one from the other
|
||
|
||
2. let t be an upper bound on the number of possible failing processes and f the real number of process failed (hence, $0\leq f\leq t\leq n-1$, with f unknown and t known in advance).
|
||
Then, there are at least $t-f$ correct processes different from $p_L$ with a timer s.t. $\exists$ time $\tau_{2} \forall$ time interval $\delta$, if their timer is set to $\delta$ after $\tau_{2}$, it expires at least after $\delta$.
|
||
|
||
REMARK: $\tau_{1}, \tau_{2}, \nabla$ and $p_L$ are all unknown.
|
||
|
||
IDEA:
|
||
- `PROGRESS[1..n]` is an array of SWMR atomic registers used by proc’s to signal that they’re alive
|
||
- pi suspects pj if pi doesn’t see any progress of pj after a proper time interval (to be guessed) set in its timer
|
||
- the leader is the least suspected process, or the one with smallest/biggest ID among the least suspected ones (if there are more than one)
|
||
- this changes in time, but not forever
|
||
|
||
Guessing the time duration for suspecting a process:
|
||
- SUSPECT[i,j] = #times pi has suspected pj
|
||
- For all k, take the t+1 minimum values in SUSPECT[1..n , k]
|
||
- Sum them, to obtain Sk
|
||
- The interval to use in the timers is the minimum Sk
|
||
- it can be proved that this eventually becomes ≥ $\nabla$
|
||
|
||
#### From obstruction-freedom to wait-freedom
|
||
**Eventually perfect:** failure detector ♢P provides each process $p_i$ a local variable $suspected_i$ such that
|
||
1. *(Eventual completeness)* eventually, $suspended_{i}$ contains all the indexes of crashed processes, for all correct $p_i$
|
||
2. (*Eventual accuracy*) eventually, $suspected_{i}$ contains only indexes of crashed processes, for all correct $p_{i}$.
|
||
|
||
**Definition:** A failure detector FD1 is **stronger** than a failure detector FD2 if there exists an algorithm that builds FD2 from instances of FD1 and atomic R/W registers.
|
||
|
||
**Proposition:** ♢P is stronger than $\Omega_{X}$
|
||
*Proof:*
|
||
Forall i
|
||
- i ∉ X $\to$ `ev_leader_i(X)` is any ID (and may change in time)
|
||
- $i \in X \to$ `ev_leader_i(X)` $= min\left( suspected_{i} \cap X \right)$ where $Π$ denotes the set of all proc. IDs.
|
||
|