Ray: A Distributed Framework for Emerging AI Applications

• p. 15 Section 7, API paragraph

  DDPG [55]

• p. 17 Reference [55]

  Deterministic policy gradient algo-rithms. In ICML (2014).

• p. 15 Section 7 Discussion, API paragraph

  DDPG [55], and Ape-X [27]

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: that DDPG was introduced by Lillicrap et al. (2015), not by Silver et al. (2014). The crops confirm that the text and reference entry read as quoted. DPG is the precursor, so the citation may be deliberate.

Second check: another model agreed

• p. 3 Section 1 Introduction, left column

  Bulk-synchronous parallel systems such as Map- Reduce [20], Apache Spark [64], and Dryad [28]

• p. 14 Section 6 Related Work, Dataflow systems

  Dryad relaxes this restriction but lacks support for dynamic task graphs.

Judge's note: Both passages checked on the crops. The two descriptions of Dryad's execution model don't fully agree. The introduction may be using 'bulk-synchronous' loosely, so this is a minor consistency point rather than an error.

Second check: another model agreed

• p. 6 Figure 3 code, Simulator class

  def rollout(self, policy, num_steps):

• p. 6 Figure 3 code, train_policy

  rollout_ids = [s.rollout.remote(policy_id)

Judge's note: Checked on the page 6 image. The example code doesn't match itself. It is an illustrative listing, so this is minor.

Second check: another model agreed

• p. 8 Section 4.2.3, right column

  we use the former to reconstruct objects in the store.

• p. 6 Section 3.2, end

  whether produced by remote functions or actor methods (Section 4.2.3).

Only the authors can say what is right here.

Judge's note: Read both passages on the page images. Section 5.1 (p.11) also says actor method calls encoded as stateful edges let Ray 'reuse the same object reconstruction mechanism'. Read literally, 'the former' limits object reconstruction to task lineage, which conflicts with the sentence that cites 4.2.3. The authors may only mean that actor-state recovery is handled separately, so this is put to them as a query.

Second check: another model agreed

• p. 12 Section 5.2.3 text

  up to 1.8× throughput

• p. 12 Table 4, 256 CPUs column

  4.03M

Judge's note: Table 4 values checked on the page 12 image. 1.8× is the ratio truncated rather than rounded, which is conservative. This is a minor consistency point only.

Second check: another model agreed

• p. 12 5.2.1, text

  is within 10% of distributed TensorFlow

• p. 12 Figure 13

  Distributed TF

Only the authors can say what is right here.

Judge's note: Bar heights read by eye from the enlarged Figure 13 crop: at 16 GPUs about 2100 vs about 1820, and at 32 GPUs about 3850 vs about 3350. Both gaps are above 10%. At 4, 8 and 64 GPUs the gaps are within 10%. Because the values are read from a plot, this is a query to the authors.

Second check: another model agreed

• p. 4 Figure 2 pseudocode

  rollout(policy, environment):

• p. 4 Section 2, left column

  ing rollout(environment, policy) to generate trajectories.

• p. 4 Section 2, text

  invok- ing rollout(environment, policy)

• p. 4 Section 2, paragraph after Figure 2

  invok- ing rollout(environment, policy) to generate trajectories

Judge's note: Page 4 image: Figure 2 has 'rollout(policy, environment):' and 'trajectories.append(rollout(policy, environment))'. The left-column prose has rollout(environment, policy) in the same bold/italic style, so the arguments are swapped.

Second check: none (judge only)

• p. 10 Figure 10b x-axis label

  Elasped Time (seconds)

Judge's note: The misspelling is visible in the rendered figure in the crop.

Second check: none (judge only)

• p. 11 Figure 12b x-axis label

  Added scheduler latency (ms

Judge's note: In the crop and on the full page 11 image, the label ends at 'ms' inside the figure area. It is not cut off by the crop.

Second check: none (judge only)

• p. 16 Reference [21]

  Proceedings of the 2Nd An-

Judge's note: Looked at the crop of p.16. Reference [21] shows the italic text 'Proceedings of the 2Nd An-nual Symposium on Computer Architecture'. The capital N comes from BibTeX title-casing and is printed in the PDF, so it is not an extraction artefact. It is cosmetic.

Second check: none (judge only)

• p. 18 Reference [61]

  Opearting Systems Design

• p. 18 Reference [61]

  6th Conference on Symposium on Opearting Systems Design

• p. 18 Reference [61]

  Symposium on Opearting Systems Design

Judge's note: Looked at the crop of p.18. Reference [61] (van Renesse and Schneider) prints 'Opearting' in the italic venue name, so the misspelling is in the typeset PDF. All three reviewer notes describe this same defect.

Second check: none (judge only)