pFabric: Minimal Near-Optimal Datacenter Transport

• p. 1 §1 Introduction, switch bullet

  if the incoming packet has lower priority than all buffered packets, it is dropped. Else, the low-est priority packet in the buffer is dropped

• p. 3 §4.1 Switch Design, Priority dropping

  if it has priority less than or equal to the lowest priority packet in the buffer, it is dropped

Judge's note: Checked both crops. The intro uses a strict comparison and §4.1 uses ≤, so the two rules drop different packets only when the arriving packet ties the lowest buffered priority. The intro is a summary and §4.1 is the precise design, so this is a minor consistency point.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 2 Introduction, first bullet

  flows compared to PDQ and DCTCP by more than 40% and

• p. 8 §5.4.1, FCT breakdown

  erage FCT for the small flows with pFabric is ∼30-50% lower for

• p. 2 §1 Introduction, first findings bullet

  pFabric reduces the FCT for short flows compared to PDQ and DCTCP by more than 40%

• p. 8 §5.4.1, FCT breakdown paragraph

  the av-erage FCT for the small flows with pFabric is ∼30-50% lower for the web search workload

Only the authors can say what is right here.

Judge's note: Checked the intro bullet on p2 and §5.4.1 and Fig. 8(a) on p8. The paper's own text gives a ~30% lower bound for web search, which is below the intro's "more than 40%". Fig. 8(a) at load 0.1 reads PDQ ≈1.4 and pFabric ≈1.05; these are approximate plot readings. The intro gives no load or workload qualifier that would reconcile the two. The authors should say what the 40% refers to.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 2 Introduction, first bullet

  2.5–4× respectively at the mean

• p. 8 Figures 8(a), 9(a)

  (a) (0, 100KB]: Avg

• p. 2 Introduction, first findings bullet

  by more than 40% and 2.5–4× respectively at the mean

• p. 8 Sec 5.4.1 FCT breakdown

  is ∼30-50% lower for the web search workload

Only the authors can say what is right here.

Judge's note: Read from the plots, so values are approximate. Fig. 8(a) at 0.8: DCTCP ≈3.6, pFabric ≈1.1. Fig. 9(a) at 0.8: DCTCP ≈2.8, pFabric ≈1.1. At 0.1 the DCTCP curve is ≈1.3–1.4. Neither plot reaches 4×, and the intro states no load qualifier. Only the authors can say which statistic gives 2.5–4×. Note 3 duplicates C002. Note 4 (the "(0, 100KB]" brackets) is not a defect: it is standard half-open interval notation.

Second check: another model agreed, code check

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 §5.4.2, right column

  ulations) from 100µs to 100ms to explore the behavior under tight

• p. 9 Figure 10 x-axis

  Average Deadline (us)

• p. 8 Sec 5.4.2

  from 100µs to 100ms

• p. 8 5.4.2

  from 100µs to 100ms to explore the behavior under tight

Only the authors can say what is right here.

Judge's note: Checked Fig. 10 on p9: both panels end at 10^4 µs. The text gives no indication that larger values were run but left off the plot. The authors should say whether 100ms should read 10ms or whether points are missing from the figure.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 11 §7 Discussion, Starvation & Gaming

  This is consistent with our findings (e.g., Fig-ures 8(c) and 9(c))

• p. 8 Figure 8(c) and §5.4.1 text

  pFabric is roughly the same as TCP and ∼30% worse than Ideal at 80% load for the large flows

Only the authors can say what is right here.

Judge's note: I looked at Fig. 8(c) on the page 8 image. pFabric (red) and TCP-DropTail (black) nearly coincide at every load and both end near 13 at load 0.8. DCTCP is lower, at about 10–11. Fig. 9(c) does show pFabric below TCP for large flows. Citing Fig. 8(c) as support for an improvement in large flows over fair sharing overstates what that figure shows. 'Consistent with' could be read loosely as 'large flows are not starved', so this is a wording query for the authors, not a data error.

Second check: another model agreed

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 3 Sec 3

  (and its associate delays)

Judge's note: The page image shows "associate delays" in §3.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 7 §5.4.1, end

  normalized FCT s 50% lower than PDQ.

• p. 7 Sec 5.4.1

  normalized FCT s 50% lower than PDQ

Judge's note: The page image shows "FCT s 50%", so this is not an extraction artefact.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 7 §5.3 Seamless switching

  This is only 150µs larger then the best possible completion time of

• p. 7 5.3 Seamless switching

  150µs larger then the best possible completion time

Judge's note: Visible in the page image of §5.3.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 7 §5.3 Incast

  ideal request complete time of 80ms across all number of senders.

Judge's note: The page image shows "complete time"; the same paragraph elsewhere uses "total request completion time".

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 §5.4.2

  Figure 10 shows the application throughout for the two work-

Judge's note: Visible in the page image; the metric is called Application Throughput elsewhere.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 8 Sec 5.4.1 FCT breakdown

  provided in [6]) The results are shown

• p. 8 Sec 5.4.1 FCT breakdown

  for the small flows The results show

Judge's note: Both missing periods are visible in the page image of §5.4.1.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 10 §6, end of intro

  time. The are two key challenges with this approach

• p. 10 Sec 6

  The are two key challenges with this approach

Judge's note: The page 10 image shows 'The are two key challenges'. This is a typo.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 11 §7 Starvation & Gaming

  cap the priority numbers so that beyond a certain size, all flows get

• p. 11 Discussion, Starvation & Gaming

  place a cap the priority numbers

Judge's note: The page 11 image shows the wording as quoted. A word is missing.

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send

• p. 12 References [21]

  A. Rowtron

Only the authors can say what is right here.

Judge's note: Rests on outside knowledge: the coauthor of Wilson et al., 'Better never than late' (SIGCOMM 2011) is Antony Rowstron. The page 12 image shows the reference as quoted, 'C. Wilson, H. Ballani, T. Karagiannis, and A. Rowtron.'

Second check: none (judge only)

Report a problem with this finding

This finding is

wrong right unclear or overstated already corrected (erratum, later version)

I am an author of this paper Send