The Tail at Scale
Communications of the ACM, 2013 · paper · Computer science
Outlines common causes of occasional high latency in large online services and techniques that tolerate it. The authors argue that such techniques can build predictably responsive systems from less predictable parts, often using resources already deployed for fault tolerance.
Checked: PDF from barroso.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label References 1 ("Barroso, L.A. and Höelzle, U. The case for energy proportional computing.") and 11 ("Mudge, T. and Hölzle, U. Challenges and opportunities for extremely energy-efficient processors.") are listed but never cited in the text.
- p. 7 References, entries 1 and 11
Barroso, L.A. and Höelzle, U. The case for energy proportional computing.
- p. 7 References, entry 11
Mudge, T. and Hölzle, U. Challenges and opportunities for extremely energy-efficient processors.
- p. 3 Energy management paragraph (no citation marker)
Power-saving modes in many types of devices save considerable energy
- p. 7 References 1 and 2
Barroso, L.A. and Höelzle, U.
Only the authors can say what is right here.
Judge's note: The text layer gives the in-text superscripts as 3, 2, 7 (p.1), 4, 5 (p.3), 10 (p.4), 10, 9 (p.5), 12, 6 (p.6) and 8 (p.7). I checked the page images. Page 1 has only 3, 2 and 7. Page 2 is a full-page illustration with no text. The Energy management bullet on p.3 ends 'active modes.' with no marker. The Hardware Trends paragraph on p.7 ('more aggressive power optimizations becoming available') has no marker. No superscript 1 or 11 appears anywhere. Only the authors can say where these citations belong. The Höelzle/Hölzle spelling point raised in this cluster is handled under C010.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference 10 (PDF p.7) gives the journal as "IEEE Transactions on Parallel and Distributed Computing 12, 10 (Oct. 2001)". The journal is presumably IEEE Transactions on Parallel and Distributed Systems.
- p. 7 References, entry 10
IEEE Transactions on Parallel and Distributed Computing 12, 10 (Oct. 2001)
- p. 7 Reference 10
IEEE Transactions on Parallel and Distributed Computing
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that Mitzenmacher's 'The power of two choices in randomized load balancing' (vol. 12, no. 10, 2001, pp. 1094-1104) appeared in IEEE Transactions on Parallel and Distributed Systems. The crop of reference 10 on p.7 confirms the printed title 'IEEE Transactions on Parallel and Distributed Computing'.
Second check: another model agreed
Report a problem with this finding
10 minor (typos and the like; most checked by the judge only)
Reference or label Reference 2 (PDF p.7) gives the series as "Synthesis Series on Computer Architecture". The Morgan & Claypool series is presumably 'Synthesis Lectures on Computer Architecture'.
- p. 7 References, entry 2
Synthesis Series on Computer Architecture
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that 'The Datacenter as a Computer' (2009) appeared in the series 'Synthesis Lectures on Computer Architecture'. The crop of p.7 confirms the printed 'Synthesis Series on Computer Architecture'. The work can still be identified, so this is minor.
Second check: another model agreed
Report a problem with this finding
Wording PDF p.3: "Techniques that concentrate on these slow outliers can yield dramatic reductions in overall service performance." Read literally, this says performance gets worse. The intended meaning is reductions in latency, or improvements in performance.
- p. 3 Component-Level Variability section, paragraph on Table 1
yield dramatic reductions in overall ser-vice performance
Judge's note: I checked the crop of p.3. The sentence reads as quoted. The preceding argument is that the slowest 5% account for half of the 99th-percentile latency (70ms of 140ms, Table 1). So the techniques are meant to reduce latency, and 'reductions in ... performance' says the opposite. A loose reading of 'performance' as response time is possible, so I treat this as a wording ambiguity with an obvious fix, not a substantive error.
Second check: another model agreed
Report a problem with this finding
Typo PDF p.3 prints "the total 99%-percentile latency". The term is redundant and does not match the paper's other forms, '99th-percentile' in the text and '99%ile' in the tables.
- p. 3 paragraph on Table 1
99%-percentile latency
Judge's note: The crop shows '99%-percentile' in print. The same paragraph uses '99th-percentile' (with a superscript th) three times, and Table 1 uses '99%ile'. This is a cosmetic inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo The x-axis of the figure on PDF p.4 is labelled "Numbers of Servers"; it should read 'Number of Servers'.
- p. 4 Figure (probability of one-second service-level response time), x-axis label
Numbers of Servers
Judge's note: The crop of the figure on p.4 shows the axis label 'Numbers of Servers' in print. This is a cosmetic grammatical slip.
Second check: none (judge only)
Report a problem with this finding
Typo PDF p.5: "the phenomena that causes variability does not tend to simultaneously affect". The plural 'phenomena' needs the plural verbs 'cause' and 'do'.
- p. 5 paragraph beginning 'Note, too, the class of techniques'
the phenomena that causes variability does not tend
- p. 5 Within-request adaptations, paragraph beginning 'Note, too'
phenomena that causes variability does not tend
Judge's note: The crop of p.5 shows the text as quoted. The next sentence treats the same word as plural ('such uncorrelated phenomena are'). This is a grammatical slip only; the meaning is clear.
Second check: none (judge only)
Report a problem with this finding
Typo PDF p.6: "as well as other to systems that inherently deal with imprecise results". The words are transposed; it should read 'as well as to other systems'.
- p. 6 Large Information Retrieval Systems, first paragraph
as well as other to systems that inher-ently deal with imprecise results
- p. 6 Large Information Retrieval Systems
as well as other to systems that inher-
Judge's note: The crop of p.6 shows 'as well as other to systems that inher-' in a single column. It is not an extraction artefact. Swapping 'other' and 'to' gives the evident fix.
Second check: none (judge only)
Report a problem with this finding
Typo PDF p.6 cites "in Stoica12 and the virtual-processor-partitioning technique in DeWitt et al.6". Reference 12 has five authors (Stoica I., Morris, R., Karger, D., Kaashoek, F., and Balakrishnan, H.) but is cited without 'et al.'.
- p. 6 Micro-partitions paragraph
in Stoica12 and the virtual-processor-partitioning technique in DeWitt et al.6
Judge's note: The crop of p.6 shows 'Stoica' followed by superscript 12, and 'DeWitt et al.' followed by superscript 6. Reference 12 on p.7 lists five authors. The reference number still resolves correctly, so this is only a cosmetic inconsistency in citation style.
Second check: none (judge only)
Report a problem with this finding
Typo The reference list spells the same co-author two ways: "Barroso, L.A. and Höelzle, U." (refs 1 and 2) versus "Mudge, T. and Hölzle, U." (ref 11).
- p. 7 References 1 and 2
Barroso, L.A. and Höelzle, U.
- p. 7 Reference 11
Mudge, T. and Hölzle, U.
- p. 7 References, entry 11
Mudge, T. and Hölzle, U. Challenges and opportunities for extremely energy-efficient processors.
- p. 7 Hardware Trends and Their Effects (no citation)
Variability at the hardware level is likely to be higher in the future
Judge's note: The p.7 crops confirm 'Höelzle' in refs 1 and 2 and 'Hölzle' in ref 11. The inconsistency is visible within the reference list itself. 'Höelzle' combines the umlaut with the 'oe' transliteration. The point in this cluster that ref 11 is uncited is covered by C009.
Second check: none (judge only)
Report a problem with this finding
Typo References 4 and 12 drop the comma after the surname: "Chang F., Dean J., Ghemawat, S." and "Stoica I., Morris, R.". Every other entry uses the 'Surname, I.' format.
- p. 7 References 4 and 12
Chang F., Dean J., Ghemawat, S.
- p. 7 Reference 12
Stoica I., Morris, R.
Judge's note: The crop of p.7 confirms that the commas are missing after Chang, Dean and Stoica, while every other author in the list has a comma after the surname. This is a cosmetic formatting inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 9 (PDF p.7) prints the journal as "SIGOPS Operating System Review 38, 5 (Oct. 2004)". The title is presumably 'Operating Systems Review'.
- p. 7 Reference 9
SIGOPS Operating System Review
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the ACM SIGOPS journal is titled 'Operating Systems Review' (plural). The crop of p.7 confirms the printed 'SIGOPS Operating System Review'. The difference is one letter and the citation can still be identified.
Second check: none (judge only)