Large-scale Cluster Management at Google with Borg
European Conference on Computer Systems, 2015 · paper · Computer science
Describes Borg, Google's cluster manager running hundreds of thousands of jobs across clusters of up to tens of thousands of machines. It covers the architecture, a quantitative analysis of some policy decisions, and lessons from a decade of operation.
Checked: PDF from static.googleusercontent.com, fetched 4 Oct 2026. Page numbers are those of that PDF.
4 findings
Reference or label Figure 10's y-axis reads 'Percentage of clusters', but the caption says the CDF is 'for 15 representative cells' and the analogous CDFs (Figs 5b, 9) are labelled 'Percentage of cells'.
- p. 9 Figure 10 y-axis label
Percentage of clusters
- p. 9 Figure 10 caption
if we disabled it for 15 representative cells
- p. 9 Figure 10 caption
for 15 representative cells
Judge's note: Page 9 image: the Fig 10 axis label says clusters and Fig 9 says cells. Page 7: Fig 5b says cells. The caption counts cells. This is an axis label inconsistent with the population.
Second check: another model agreed
Report a problem with this finding
Reference or label The Figure 13 caption says 'did a thread have to wait longer than 5 ms to access a CPU (the white bars)', but the legend defines the white bars as '[1ms, 5ms)'; the darker bars are [5ms, 10ms) and [10ms, inf).
- p. 10 Figure 13 caption
In only a few percent of the time did a thread have to wait longer than 5 ms to access a CPU (the white bars)
- p. 10 Figure 13 legend
[1ms, 5ms)
- p. 10 Figure 13 caption vs legend
longer than 5 ms to access a CPU (the white bars)
Judge's note: Page 10 image: the legend has white = [1ms, 5ms), grey = [5ms, 10ms) and black = [10ms, inf). The caption opens with 'wait longer than 1ms' and says 'they almost never had to wait longer (the darker bars)', so the white-bar threshold should be 1 ms. The bracket notation flagged by reviewer note 3 is a deliberate half-open interval and is not a defect.
Second check: another model agreed, code check
Report a problem with this finding
Reference or label Ref [22] reads "Communications of the ACM, 56(2):74–80, Feb. 2012."; volume 56 issue 2 of CACM is dated Feb. 2013.
- p. 15 Reference [22]
Communications of the ACM, 56(2):74–80, Feb. 2012.
- p. 15 Reference [22]
56(2):74–80, Feb. 2012
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: CACM volume 56 is the 2013 volume and 'The tail at scale' appeared in Feb. 2013. The crop of p.15 shows the entry exactly as quoted. Ref [23] on the same page gives CACM 51(1) as 2008. With one volume per year, that fits volume 56 being 2013, but the paper does not state the mapping, so I cannot confirm it from the files alone.
Second check: another model agreed
Report a problem with this finding
Reference or label Ref [35] (Dominant Resource Fairness, NSDI 2011) gives "pages 323–326, 2011.", a 4-page range for a full conference paper; the proceedings range is 323–336.
- p. 16 Reference [35]
(NSDI), pages 323–326, 2011.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the DRF paper occupies pages 323–336 of the NSDI 2011 proceedings. The crop of p.16 shows "(NSDI), pages 323–326, 2011." The printed range covers only 4 pages, which looks like a digit error, but the files cannot confirm the true range.
Second check: another model agreed
Report a problem with this finding
13 minor (typos and the like; most checked by the judge only)
Wording §2.5 defines priority as 'a small positive integer' but later says 'every user has infinite quota at priority zero'; zero is not positive.
- p. 3 Section 2.5, para 2
Every job has a priority, a small positive integer.
- p. 3 Section 2.5, quota paragraph
every user has infinite quota at priority zero
Judge's note: Both sentences are in §2.5 on p.3 (crops a and b). Zero is used as an ordinary priority level, so the definition is inexact. 'Non-negative' would make the two consistent. Wording-level only.
Second check: another model agreed
Report a problem with this finding
Wording §5.2 says 'even with the larger threshold, we would need 2–16× as many cells, and 20–150% additional machines', but in Figure 6 the 100 TiB points span only about 2–3 cells and about 25–50% overhead; the upper ends belong to the 10 TiB points.
- p. 8 Section 5.2, Figure 6 discussion
even with the larger threshold, we would need 2–16× as many cells, and 20–150% additional machines
- p. 7 Figure 6
100 TiB RAM/user
Only the authors can say what is right here.
Judge's note: Figure 6 (p.7) read from the image: the blue 100 TiB points sit at about 2–3 cells with about 25–50% overhead. The orange 10 TiB points sit at about 8–16 cells with about 70–155% overhead. The quoted ranges therefore span both thresholds, not the larger one alone. The highest orange point looks slightly above 150%, but that reading is approximate. The authors should say whether the sentence means across both thresholds.
Second check: another model agreed
Report a problem with this finding
Typo §2.1 spells the system 'BigTable' (p.1, '(e.g., BigTable)') and then 'Bigtable [19]' (p.2), matching ref [19] 'Bigtable'.
- p. 1 Section 2.1
BigTable
- p. 2 Section 2.1, para 3
Bigtable [19]
Judge's note: Both spellings appear in the page images (p.1 right column, p.2 left column). The reference title uses 'Bigtable'. Cosmetic inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo The text has 'Millwheel [3]' (p.2), but reference [3] is titled 'MillWheel: fault-tolerant stream processing'.
- p. 2 Section 2.1, para 3
Millwheel [3]
- p. 15 Reference [3]
MillWheel: fault-tolerant stream processing
Judge's note: I confirmed both forms in the p.2 and p.15 crops. Capitalisation inconsistency only.
Second check: none (judge only)
Report a problem with this finding
Typo In §3.4 the cross-reference has no parentheses: 'across the five Borgmaster replicas §3.3.'; elsewhere the paper uses e.g. '(§3.4)'.
- p. 5 Section 3.4, last paragraph
across the five Borgmaster replicas §3.3.
Judge's note: Page 5 image, end of the right column. The target is plausible: §3.3 describes link shards running on each Borgmaster replica. Only the parenthesisation is inconsistent, which is cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo The Figure 11 caption has a subject–verb agreement error: 'The dotted lines shows CDFs of the ratio'.
- p. 9 Figure 11 caption
The dotted lines shows CDFs of the ratio
- p. 9 Figure 11 caption
The dotted lines shows CDFs
Judge's note: Visible in the page 9 image. The parallel sentence later reads 'The solid lines show'.
Second check: none (judge only)
Report a problem with this finding
Typo Wrong article in §5.5: 'and then to an medium setting' should read 'a medium setting'.
- p. 10 Section 5.5, last paragraph
and then to an medium setting
- p. 10 Sec 5.5, last paragraph
to an medium setting
- p. 10
and then to an medium setting that was mid-way between
Judge's note: Visible in the page 10 image, end of the right column.
Second check: none (judge only)
Report a problem with this finding
Typo The text 'Isaard [48] outlines' and ref [48] 'M. Isaard. Autopilot' spell the name differently from ref [49] 'M. Isard, V. Prabhakaran, ...'.
- p. 12 Section 7, Autopilot paragraph
Isaard [48] outlines many best practices
- p. 16 Reference [48]
M. Isaard. Autopilot: Automatic data center management.
- p. 16 Reference [49]
M. Isard, V. Prabhakaran, J. Currey
- p. 12 Related work, Autopilot paragraph
Isaard [48] outlines
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the Autopilot author is Michael Isard, the same person as in [49]. The page 12 and page 16 images confirm 'Isaard' in the text and in [48], and 'Isard' in [49]. Within the paper this is a spelling inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Stray comma between subject and verb in §7: 'Facebook’s Tupperware [64], is a Borg-like system'.
- p. 12 §7 Related work, Tupperware paragraph
Facebook’s Tupperware [64], is a Borg-like system
Judge's note: Crop of page 12 shows the comma, and no matching comma follows. Punctuation slip.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [24] (Paragon, ASPLOS) has a truncated year: '(ASPLOS), Mar. 201.'
- p. 15 Reference [24]
and Operating Systems (ASPLOS), Mar. 201.
- p. 15 Reference [24]
Mar. 201.
- p. 15 Reference [24]
(ASPLOS), Mar. 201.
Judge's note: The page 15 crop shows 'Mar. 201.' in the image itself, so it is not an extraction artefact. The truncation is evident. The intended year (2013) rests on outside knowledge.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [11] spells the author as "U. Holzle", while ref [10] spells the same author "U. Hölzle".
- p. 15 Reference [11]
L. A. Barroso, J. Dean, and U. Holzle.
- p. 15 Reference [10]
U. Hölzle
Judge's note: The crop of p.15 shows [10] with the umlaut ("U. Hölzle") and [11] without it ("U. Holzle"). The images show this, so it is not an extraction artefact. It is an internal spelling inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [59] gives the venue as "Indianapolis, IA, USA, 2010"; the postal abbreviation for Indiana is IN, not IA (Iowa).
- p. 16 Reference [59]
Indianapolis, IA, USA, 2010
- p. 16 Reference [65]
Farminton, PA
- p. 16 Reference [59]
Indianapolis, IA
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: Indianapolis is in Indiana (IN), and IA is the abbreviation for Iowa. The crop of p.16 shows "pages 135–146, Indianapolis, IA, USA, 2010." One reviewer's note also mentions the "Farminton" misspelling in ref [65]; that is handled separately as C019.
Second check: none (judge only)
Report a problem with this finding
Typo Ref [65] gives the venue as "69–84, Farminton, PA, USA, 2013."; the city is spelled Farmington.
- p. 16 Reference [65]
69–84, Farminton, PA, USA, 2013.
- p. 16 Reference [65]
Farminton, PA, USA
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: SOSP 2013 was held in Farmington, PA. The crop of p.16 shows "Farminton" without the 'g', so it is not an extraction artefact.
Second check: none (judge only)