test_posture¶
Resolving a posture into a share, and delivering that share.
Attributes¶
Functions¶
The definitional postures are the point of offering them: a caller who |
|
The share is where both finish together, so neither waits. |
|
Constrained against the MEASUREMENT, not against the formula restated. |
|
The reason to check before building any of this: at parity a split |
|
|
A host cost may be measured, but it is never guessed: with neither a |
The measurement closes the loop the closed form leaves open: a caller |
|
A silent fall back to collaborative would answer a different question |
|
Recorded because it is counterintuitive and a planner that only |
|
|
|
No posture may be silently absent -- a name in POSTURES that nothing |
|
With no transfer to hide, every batch is pure per-call overhead. |
|
45.8 MiB measured fastest at 3 batches (18.78 ms), against 22.08 at one |
|
|
Which is the entire reason the batch count depends on residency: a |
Splitting below the knee costs transfer rate for a tail that is already |
|
Rounding up contradicted the measurement this targets: at 1.5x the knee |
|
|
It is a measurement of one link, and a library-wide constant would be |
|
A device job costs something before its first row. The balance point |
Module Contents¶
- test_posture.KNEE_BYTES = 15728640¶
- test_posture.test_cpu_and_offload_need_no_estimates() None¶
The definitional postures are the point of offering them: a caller who will not pay for cost estimation can still state an intent.
- test_posture.test_collaborative_balances_the_engines() None¶
The share is where both finish together, so neither waits.
- test_posture.test_the_closed_form_lands_on_a_measured_optimum() None¶
Constrained against the MEASUREMENT, not against the formula restated.
Asserting the formula’s own output would pass against any formula, so the assertions are that the predicted SHARE lands on the measured optimum and that the predicted speedup is within the error this model claims.
- test_posture.test_a_split_pays_least_when_one_engine_dominates() None¶
The reason to check before building any of this: at parity a split returns 2x, and against a ten-times-faster engine it buys 1.1x over simply handing the whole query over.
- test_posture.test_collaborative_refuses_rather_than_inventing_a_cost_model() None¶
A host cost may be measured, but it is never guessed: with neither a figure nor a profile, a fabricated number would sit underneath a real decision.
- test_posture.test_collaborative_rejects_nonpositive_estimates() None¶
- test_posture.test_a_measured_profile_supplies_the_host_figure() None¶
The measurement closes the loop the closed form leaves open: a caller that profiled its host does not also have to price it by hand.
- test_posture.test_a_profile_prices_a_row_count() None¶
- test_posture.test_adaptive_refuses_instead_of_falling_back() None¶
A silent fall back to collaborative would answer a different question than the caller asked – adaptive chooses a boundary in the plan, not a row share.
- test_posture.test_offload_is_not_the_wall_clock_optimum() None¶
Recorded because it is counterintuitive and a planner that only minimized wall time would pick wrongly: on the measured pairing offload reaches 1.49x where collaborative reaches 2.49x, and buys back the host.
- test_posture.test_unknown_posture_is_refused() None¶
- test_posture.test_split_validates_its_own_fields() None¶
- test_posture.test_every_posture_is_reachable_or_refused_by_name() None¶
No posture may be silently absent – a name in POSTURES that nothing handles would fall through to the unknown-posture branch and read as a typo rather than a gap.
adaptiveis refused by name, as a decision this signature cannot make, and the refusal must not be the unknown-posture one.
- test_posture.test_resident_wants_one_batch_at_any_size() None¶
With no transfer to hide, every batch is pure per-call overhead. Measured: 5.27 ms at one batch against 5.91 at three.
- test_posture.test_streamed_reproduces_the_measured_optimum() None¶
45.8 MiB measured fastest at 3 batches (18.78 ms), against 22.08 at one and 22.59 at twelve. Targeting the link’s knee lands there.
- test_posture.test_the_two_optima_are_opposite() None¶
Which is the entire reason the batch count depends on residency: a single tuned constant would be wrong for one of them.
- test_posture.test_small_payloads_are_not_split_below_the_knee() None¶
Splitting below the knee costs transfer rate for a tail that is already short.
- test_posture.test_batches_are_never_split_below_the_knee() None¶
Rounding up contradicted the measurement this targets: at 1.5x the knee it produced two batches of 0.75x each, in the degraded region.
- test_posture.test_batch_count_rejects_impossible_inputs() None¶
- test_posture.test_the_knee_has_no_default() None¶
It is a measurement of one link, and a library-wide constant would be wrong for every link that is not the one it was measured on.
- test_posture.test_the_device_fixed_cost_moves_the_balance_and_can_rule_the_split_out() None¶
A device job costs something before its first row. The balance point hands the device less by exactly that much, the predicted speedup is against the wall that includes it, and when the fixed cost is most of the host’s whole time no share can win and the answer is host-only.