Arrakis: The Operating System is the Control Plane
2013 · paper · Computer science
Presents Arrakis, an operating system that lets applications access virtualized I/O devices directly, so the kernel works only in the control plane. For a NoSQL store, the authors report 2-5× lower latency and 9× higher throughput than well-tuned Linux.
Checked: PDF from usenix.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
5 findings
Substance §4.4 says Redis "write latency improves by 63%", but Table 2 (163.14 → 31.51 µs, an 80.7% reduction, 5.2×) and the abstract, introduction ("5× better write latency") and conclusion ("2–5× faster") imply about 81%.
- p. 12 §4.4 Redis NoSQL Store, 2nd column
write latency improves by 63%
- p. 2 Introduction
5× better write la-
- p. 5 Table 2, Durable write totals
163.14
- p. 15 Conclusion
store is 2–5× faster
Only the authors can say what is right here.
Judge's note: Checked page 12 §4.4, Table 2 on page 5, the introduction on page 2 and the conclusion on page 15. No quantity in the paper gives 63%: the mean is 80.7% and the 99th percentile is 81.0%. A 63% reduction corresponds to 2.7×. The conclusion calls the 5× figure end-to-end client latency, so the two statements conflict. Only the authors can say which is right.
Second check: another model agreed
Report a problem with this finding
Substance Table 2, Read hit Arrakis/P column: the components (1.12, 0.29, 0.66, 0.10, 0.64, 0.71, 0.46) sum to 3.98 µs, not the printed Total "4.07", and the printed percentages sum to 97.79%.
- p. 5 Table 2, Read hit / Arrakis/P column
4.07
- p. 5 Table 2, Read hit, Arrakis/P column
4.07 (σ = 0.44)
Only the authors can say what is right here.
Judge's note: Checked Table 2 on the page 5 image. The other three columns (Linux read 8.67, Linux write 163.14, Arrakis write 31.51) sum exactly to their totals, including the 'Other' row. Rounding of seven two-decimal entries cannot explain a 0.09 µs gap. Each percentage is computed against 4.07 (e.g. 1.12/4.07 = 27.52%), so a component appears to be misprinted or missing. Only the authors can say which.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [36] lists "M. Jamshed, H. J. S. Ihm, D. Han", apparently merging two authors (H. Jeong and S. Ihm) into one.
- p. 17 Reference [36]
M. Jamshed, H. J. S. Ihm, D. Han
- p. 17 Reference [36]
M. Jamshed, H. J. S. Ihm, D. Han,
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the mTCP (NSDI 2014) author list includes H. Jeong and S. Ihm as separate authors. The page 17 image shows 'H. J. S. Ihm' as printed.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [39]: "IEEE J.Sel. A. Commun., 14(7):1280–1297, Sep 2006." The year appears to be wrong; volume 14 of JSAC is from 1996.
- p. 17 Reference [39]
14(7):1280–1297, Sep 2006
- p. 17 Reference [39]
IEEE J.Sel. A. Commun., 14(7):1280–1297, Sep 2006.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: IEEE JSAC volume 14, issue 7 appeared in September 1996, and the Nemesis paper dates from 1996. The page 17 image shows 'Sep 2006' as printed.
Second check: another model agreed
Report a problem with this finding
Wording §3.6: "This stack is designed to maximize both latency and throughput." The intended meaning is minimize latency and maximize throughput.
- p. 9 §3.6, first paragraph
This stack is designed to maximize both latency and throughput
- p. 9 Sec. 3.6
This stack is designed to maximize both
- p. 9 Section 3.6
This stack is designed to maximize both latency and throughput.
Judge's note: The page 9 image confirms the wording. Taken literally, the sentence says the opposite of the paper's design goal (low latency). The fix is obvious.
Second check: another model agreed
Report a problem with this finding
6 minor (typos and the like; most checked by the judge only)
Wording §2.1 attributes the residual idle-case cost to "a further 0.6 µs slowdown in other parts of the network stack", but Table 1's 'Network stack' rows (in/out) grow by only 0.35 µs; the 0.6 µs is the residual over network stack + copy + kernel crossing.
- p. 3 §2.1, last paragraph of left column
a further 0.6 µs slowdown in other parts of the network stack
- p. 3 Table 1, Network stack in/out rows
1.24
Judge's note: Checked Table 1 and the §2.1 text on page 3. Total rises from 3.36 to 6.19 µs; the scheduler accounts for 2.23 (≈2.2 µs as printed). The remaining 0.60 µs is spread over Network stack (+0.35), Copy (+0.12) and Kernel crossing (+0.13). Table 1 defines 'Network stack' as its own category, so the prose uses the term more loosely than the table. Either reading could be intended, so this is a minor wording point.
Second check: another model agreed
Report a problem with this finding
Wording §6 names reference [26] "Exitless IPIs [26]", while [26] is "ELI: bare-metal performance for I/O virtualization" and the sentence itself describes demultiplexing hardware interrupts, not inter-processor interrupts.
- p. 15 §6 Related Work, 3rd paragraph
Exitless IPIs [26]
- p. 16 Reference [26]
ELI: bare-metal performance for I/O virtualization
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that ELI stands for Exit-Less Interrupts and concerns device interrupts rather than IPIs. Within the paper, the sentence's own description ("demultiplex hardware interrupts between virtual machines") does not fit the label 'IPIs'. Checked page 15 and reference [26] on page 16.
Second check: another model agreed
Report a problem with this finding
Typo §3.4: "then be assigned to a specific queue on a VNIC" is typeset with 'assign' in bold and 'ed' in roman, as though it were an API identifier.
- p. 8 §3.4, paragraph on create_filter
then be assigned to a specific queue on a VNIC
Judge's note: The crop of page 8 shows the bold 'assign' with a roman 'ed'. A grep of the text finds no 'assign' API call defined anywhere in the paper. This is a cosmetic formatting slip.
Second check: none (judge only)
Report a problem with this finding
Typo §6: "for high throughout, low overhead communication" should read 'high throughput'.
- p. 15 §6 Related Work, 2nd paragraph
for high throughout, low overhead communication
- p. 15
for high throughout, low overhead communication.
Judge's note: The page 15 crop shows 'throughout' as printed. It is a misspelling.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [45]: "W. Zwanepoel" appears to be a misspelling of Zwaenepoel.
- p. 17 Reference [45]
W. Zwanepoel
- p. 17
W. Zwanepoel. IO-Lite
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the IO-Lite co-author's name is spelled Zwaenepoel. The page 17 image shows 'Zwanepoel' as printed.
Second check: none (judge only)
Report a problem with this finding
Typo Company name capitalized inconsistently in adjacent references: [50] "SolarFlare Communications, Inc." vs [51] "Solarflare Communications, Inc."
- p. 17 Reference [50]
SolarFlare Communications, Inc.
Judge's note: I checked the crop of the reference list on PDF page 17. Reference [50] prints "SolarFlare Communications, Inc. OpenOnload." Reference [51] prints "Solarflare Communications, Inc. Solarflare SFN5122F Dual-Port 10GbE Enterprise Server Adapter, 2010." The text layer only finds the [50] form, probably because the "fl" ligature in [51] breaks the search, but the image shows both forms. This is a purely cosmetic capitalization inconsistency.
Second check: none (judge only)