Tachyon: Reliable, Memory Speed Storage for Cluster Computing Frameworks
Proceedings of the ACM Symposium on Cloud Computing, 2014 · paper · Computer science
Presents Tachyon, a distributed file system for sharing data at memory speed across cluster frameworks, using lineage-based recovery instead of replication, with a checkpointing algorithm bounding recovery cost. The authors report writes 110x faster than in-memory HDFS and a realistic workflow 4x faster.
Checked: PDF from people.eecs.berkeley.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
6 findings
Substance §7.3.1 says "With the same 5000-node cluster, recomputation consumes only up to 0.9% and 1.6% of resources at Facebook and Bing respectively", but in Figure 10 these maxima are on the 1000-node curves.
- p. 11 §7.3.1 Real traces
With the same 5000-node cluster, recomputation consumes only up to 0.9% and 1.6% of resources at Facebook and Bing respectively.
- p. 12 Figure 10 plots
Facebook Workload Analysis
- p. 11 Sec 7.3.1, Real traces
With the same 5000-node cluster, recomputation consumes only up to 0.9% and 1.6% of resources at Facebook and Bing
- p. 12 Figure 10 plots and caption
recomputation consumes up to 0.9% and 1.6% of the resource in the worst case respectively
Only the authors can say what is right here.
Judge's note: I checked the Figure 10 crop on page 12. In both panels the 1000-node series (blue diamonds) is highest, at about 0.83% (Facebook) and about 1.45% (Bing) at 1024 GB. The 5000-node series is lowest, at about 0.17% and about 0.3%. The text ties the 0.9% and 1.6% maxima to 'the same 5000-node cluster' from the worst-case paragraph (5000 nodes, 1 TB each). The caption and the introduction give these values only as overall worst cases, with no cluster size. The authors need to say whether the cluster size or the values are intended.
Second check: another model agreed
Report a problem with this finding
Substance Sec 5.2 says "The rest of nodes can be recomputed when all of their children become available"; under the paper's own definition the node depends on its parents, not its children.
- p. 8 §5.2 Recomputation Order
The rest of nodes can be recomputed when all of their children become available.
- p. 8 5.2
when all of their children become available
Judge's note: Page 8, Sec 5.2, defines the relation in the same paragraph: 'The parents of a child node in the DAG denote the files that the child depends on.' The previous sentence says nodes with 'no lost parents' are recomputed first. So a remaining node can be recomputed only once its parents (its inputs) are available. Its children depend on it, so the sentence gets the dependency direction backwards.
Second check: another model agreed
Report a problem with this finding
Reference or label Figure 6 labels the recomputation jobs R2 and R3, but the example it illustrates has jobs R1 and R2 ("both R1 and R2 are launched immediately with share 1 in WR").
- p. 8 Figure 6
R3
- p. 7 §5 / §5.1 Fair Sharing
both R1 and R2 are launched immediately with share 1 in WR.
- p. 7 Figure 6 vs Sec 5.1
both R1 and R2 are launched immediately with share 1 in WR
- p. 8 Figure 6
R2 R3
Judge's note: The page 8 image shows Figure 6 with jobs R2 and R3 under WR(1) in every panel, and R3 remaining in the last panel. Page 7 defines lost files F1 and F2 with recomputation jobs R1 and R2, describes R2 finishing, and then says 'Figure 6 illustrates'. The figure labels do not match the text, which is a labelling error. Figure 5 correctly uses R2 only.
Second check: another model agreed
Report a problem with this finding
Reference or label References [17] "R. Bose and J. Frew. Lineage Retrieval for Scientic Data Processing: A Survey. In ACM Computing Surveys 2005." and [18] "R. Bose and J. Frew. Lineage retrieval for scientific data processing: a survey. ACM Computing Surveys (CSUR), 37 (1):1–28, 2005." are the same work listed twice. The text cites them separately ("scientific computing [17, 47]" and "[18] and [21] provide detailed surveys").
- p. 14 References [17], [18]
R. Bose and J. Frew. Lineage Retrieval for Scientic Data Processing: A Survey. In ACM Computing Surveys 2005.
- p. 14 Reference [17]
Lineage Retrieval for Scientic Data Processing: A Survey. In ACM Computing Surveys 2005.
- p. 14 Reference [18]
Lineage retrieval for scientific data processing: a survey. ACM Computing Surveys (CSUR), 37 (1):1–28, 2005.
- p. 12 Sec 8, Lineage paragraph
scientific comput- ing [17, 47]
Judge's note: I checked p14: both entries have the same authors, title (apart from case and the misspelling), venue and year. On p12 (Lineage paragraph) [17] and [18] are cited in separate places as if they were different works. This is a duplicated bibliography entry.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [26] (GFS) reads "In Proceedings of the ACM SIGOPS 22nd Symposium on Operating Systems Principles, 2003", but reference [13] gives the same "ACM SIGOPS 22nd Symposium on Operating Systems Principles" with "ACM, 2009", so one symposium number has two different years.
- p. 14 Reference [26]
In Proceedings of the ACM SIGOPS 22nd Symposium on Operating Systems Principles, 2003.
- p. 13 Reference [13]
Proceedings of the ACM SIGOPS 22nd Symposium on Operating Systems Principles, pages 1–14. ACM, 2009.
- p. 14 Reference [26]
Proceedings of the ACM SIGOPS 22nd Symposium on Operating Systems Principles, 2003
- p. 13 Reference [13]
Proceedings of the ACM SIGOPS 22nd Symposium on Operating Systems Principles, pages 1–14. ACM, 2009
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that GFS appeared at the 19th SOSP (2003), which would make the ordinal in [26] the wrong one. I checked both entries on the page images (p13 crop and p14). One ordinal with two years is a contradiction inside the reference list. Which entry is wrong cannot be settled from the files alone.
Second check: another model agreed
Report a problem with this finding
Wording §7.2.2 defines "Let X be the percent of jobs that output permanent data" but illustrates it with "60% (X = 60) of generated data got deleted within 16 minutes" and then sets "X as 60".
- p. 11 §7.2.2, parameter list
Let X be the percent of jobs that output permanent data. For example, 60% (X = 60) of generated data got deleted
- p. 11 Sec 7.2.2, parameter list
Let X be the percent of jobs that output permanent data. For example, 60% (X = 60) of generated data got deleted within 16 minutes
- p. 11 Sec 7.2.2
we set X as 60 and Y as 84
- p. 11 Sec 7.2.2
60% (X = 60) of generated data got deleted within 16 minutes at Facebook
Only the authors can say what is right here.
Judge's note: Page 11. Under the stated definition, 60% of data deleted would give a permanent share of about 40, not 60. The definition and the example contradict each other as printed, so it is unclear what X = 60 meant in the simulation. Only the authors can say which is intended.
Second check: another model agreed
Report a problem with this finding
28 minor (typos and the like; most checked by the judge only)
Reference or label Reference [41] gives "In 28th International Symposium on Fault-Tolerant Computing, 1997."; the ordinal and the year appear not to match.
- p. 14 Reference [41]
In 28th International Symposium on Fault-Tolerant Computing, 1997.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that FTCS-28 was held in 1998 and FTCS-27 in 1997. The page image of p14 shows the entry as quoted. Nothing in the paper allows an internal check.
Second check: another model agreed
Report a problem with this finding
Wording Contributor affiliations are described three ways: Introduction "over 15 institutions", Sec 6 "more than 15 companies and universities", Conclusion "over 15 companies".
- p. 2 Introduction
has more than 40 contributors from over 15 institutions
- p. 8 Sec 6 Implementation
has over 40 contributors from more than 15 companies and universities
- p. 13 Sec 10 Conclusion
contributions from more than 40 individuals and over 15 companies
Judge's note: I checked all three crops (pages 2, 8, 13). The Conclusion says 'companies' only, while Sec 6 counts companies and universities together. This is a wording inconsistency, minor in nature.
Second check: another model agreed
Report a problem with this finding
Wording The text says that, based on Table 3, a file is "high-priority if it has an access count higher than 2" and that "86% of the checkpointed files are leaves". Table 3 has no count-2 column; its 86% entry is under access count 3 (counts 1/3/5/10 → 62%/86%/93%/95%).
- p. 6 §4.1 Checkpointing Hot Files; Table 3
we consider a file high-priority if it has an access count higher than 2. For this workload, 86% of the checkpointed files are leaves
- p. 6 Sec 4.1 Checkpointing Hot Files / Table 3
high-priority if it has an access count higher than 2. For this workload, 86% of the checkpointed files are leaves
- p. 6 Table 3
62% 86% 93% 95%
- p. 6 Table 3 header
Access Count 1 3 5 10
Only the authors can say what is right here.
Judge's note: Page 6 image: the Table 3 header is Access Count 1, 3, 5, 10 with Percentage 62%, 86%, 93%, 95%. Integer counts 'higher than 2' means 3 or more accesses. The table gives no share for counts of 2 or fewer; it gives 86% only for 3 or fewer. The 86% leaf share of checkpointed files may be a separately measured quantity, but it matches the count-3 column, while the text says 'Based on this'. Only the authors can say whether the threshold should be '>3', or where the 86% comes from.
Second check: another model agreed
Report a problem with this finding
Wording The text says "Figure 3b shows that file requests in the aforementioned Facebook cluster is highly bursty", but Figure 3b plots data generation rates (y-axis 'Output Data Size (GB)'), not file requests.
- p. 6 §4.1 Dealing with Large Data Sets; Figure 3b
Figure 3b shows that file requests in the aforementioned Facebook cluster is highly bursty.
Judge's note: On the page 6 image, Figure 3b's caption is '(b) Data generation rates at five minutes granularity' and its y-axis is Output Data Size (GB). The text misdescribes what the figure measures, although it does show bursty writes.
Second check: another model agreed
Report a problem with this finding
Wording "we illustrate actual recomputation times when using edge caching" names the technique 'edge caching'; everywhere else the paper calls it Edge checkpointing.
- p. 7 §4.2, after Corollary 2
times when using edge caching.
Judge's note: The page 7 crop shows 'edge caching'. Sec 4 and Corollary 2 describe 'Edge checkpointing', and caching is a separate concept in the paper. This looks like a wording slip.
Second check: another model agreed
Report a problem with this finding
Wording §7.1.2 says "The pipeline ran in 16.6 minutes on Tachyon and 67 minutes on HDFS", but the baseline measured is MemHDFS (as in Figure 8 and the next sentence).
- p. 10 §7.1.2
and 67 minutes on HDFS.
Judge's note: Page 10. Figure 8's bar is labelled MemHDFS (about 67 min), and the surrounding text compares against MemHDFS throughout, so 'HDFS' names the disk-based system rather than the one measured. The ratios are consistent with the printed 4x and 3.8x.
Second check: another model agreed
Report a problem with this finding
Wording §7.1.2 states "240 jobs in 20 batches (8 Spark jobs and 4 MapReduce jobs per batch)", but the per-batch procedure describes only "two Grep applications" followed by "Word Count".
- p. 10 §7.1.2
240 jobs in 20 batches (8 Spark jobs and 4 MapReduce jobs per batch)
- p. 10 §7.1.2
we ran two Grep applications to pre-process the data. Then we ran Word Count
Only the authors can say what is right here.
Judge's note: Page 10. The total is consistent (20 × 12 = 240), but the text never explains how two Grep applications and one Word Count make up 8 Spark jobs and 4 MapReduce jobs; perhaps each application spawns several jobs. Only the authors can clarify.
Second check: another model agreed
Report a problem with this finding
Wording §7.3.1 says "recomputation takes 0.7% (655 seconds / 24 hours)", but 655/86400 = 0.76%, which rounds to 0.8%.
- p. 11 §7.3.1 Worst-case analysis
tion takes 0.7% (655 seconds / 24 hours) of the cluster’s run-
- p. 11 Sec 7.3.1 Worst-case analysis
recomputation takes 0.7% (655 seconds / 24 hours)
Judge's note: Page 11. The value appears to be truncated rather than rounded. The later 30% figure for 5000 nodes with 1 TB each (5x nodes, 8x memory) matches 0.76% (giving 30.3%) better than 0.7% (giving 28%). The 655 s holds with binary units (128×1024/200).
Second check: another model agreed
Report a problem with this finding
Typo Table 1 column header is misspelled: "Bandwith" should read "Bandwidth".
- p. 2 Table 1 header
Bandwith
- p. 2 Table 1
SDD (x4)
Judge's note: The crop of the Table 1 header on page 2 shows "Bandwith" exactly as printed, so this is not an extraction artefact. The 'SDD' row label that some reviewers also raised is judged separately under C002.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 row label "SDD (x4)" should read "SSD (x4)".
- p. 2 Table 1, row 2
SDD (x4)
Judge's note: The page 2 crop shows the row label "SDD (x4)". Everywhere else the paper writes SSD (e.g. footnote 2 'SSD- and disk-based', 'disk (or SSD)', 'The emergence of SSDs').
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 2 has a subject-verb agreement error: "This paper focus on in-memory Tachyon deployment" should read "focuses on".
- p. 2 Footnote 2
This paper focus on in-memory Tachyon deployment.
- p. 2 Footnote 2
This paper focus on in-memory Tachyon deployment
Judge's note: The page 2 crop of footnote 2 shows 'This paper focus on' as printed.
Second check: none (judge only)
Report a problem with this finding
Typo Sec 3.1 has an unclosed parenthesis: "recompute it by (partially re-executing P."
- p. 3 §3.2
recompute it by (partially3 re-executing P.
- p. 3 Sec 3.2
(partially3 re-executing P.
Judge's note: The page 3 crop shows '(partially' followed by superscript footnote mark 3, and the parenthesis is never closed before the full stop.
Second check: none (judge only)
Report a problem with this finding
Typo Grammar slips: "because of the followings:" (Sec 3.8, should be 'following'), "file requests ... is highly bursty" (should be 'are'), and "checkpoint leafs" (the paper elsewhere writes 'leaves').
- p. 4 §3.8
because of the followings:
- p. 4 3.8
because of the followings
- p. 6 Dealing with Large Data Sets
checkpoint leafs that are far apart
- p. 6 same
file requests in the aforementioned Facebook cluster is highly bursty
Judge's note: The page 4 and page 6 crops show all three phrases as printed.
Second check: none (judge only)
Report a problem with this finding
Typo Missing possessive 's' in "Tachyon’ background checkpointing"; it should read 'Tachyon’s'.
- p. 5 §4, paragraph 3
Tachyon’ background checkpointing
Judge's note: The page 5 crop shows 'Tachyon’ background checkpointing'.
Second check: none (judge only)
Report a problem with this finding
Typo "accessed less than (or equal) than some number of times" should read "(or equal to)".
- p. 6 §4.1 Checkpointing Hot Files
accessed less than (or equal) than
Judge's note: The page 6 crop shows the phrase with 'than' doubled.
Second check: none (judge only)
Report a problem with this finding
Typo Wrong article: "recomputed by jobs with a equal share under WR" should read "an equal share".
- p. 7 §5.1 Fair Sharing
with a equal share under WR
- p. 7 Fair sharing
jobs with a equal share under WR
Judge's note: The page 7 crop shows 'with a equal share'.
Second check: none (judge only)
Report a problem with this finding
Typo Subject-verb agreement: "J2 and J3 resumes all of their previous shares" should read "resume".
- p. 7 Fair sharing
J2 and J3 resumes all of their previous shares
Judge's note: Confirmed in the page 7 image (Fair Sharing paragraph).
Second check: none (judge only)
Report a problem with this finding
Typo §6 Implementation misspells 'lineage': "can take advantage of the linage transparently".
- p. 8 §6 Implementation
linage transparently
- p. 8 Section 6
take advantage of the linage transparently
- p. 8
can take advantage of the linage transparently
Judge's note: I checked the crop of page 8. The left column reads 'linage transparently', and the next line uses 'lineage' correctly.
Second check: none (judge only)
Report a problem with this finding
Typo §7.2.1 is missing a verb: "When too small an interval picked, checkpointing cannot keep up" (should read 'is picked').
- p. 10 §7.2.1
When too small an interval picked
Judge's note: Page 10, right column, confirmed in the crop.
Second check: none (judge only)
Report a problem with this finding
Typo The Figure 8 caption has a subject–verb disagreement: "In case of node failure, applications recovers in Tachyon around one minute".
- p. 10 Figure 8 caption
applications recovers in Tachyon
Judge's note: Confirmed in the Figure 8 caption on the page 10 image.
Second check: none (judge only)
Report a problem with this finding
Typo §7.3.2 has a malformed possessive: "the unimpacted users’s execution time" (should be 'user’s').
- p. 11 §7.3.2
the unimpacted users’s execution
- p. 11 7.3.2
unimpacted users’s execution
- p. 10 7.2.1
When too small an interval picked
Judge's note: Confirmed on the page 11 image. This cluster also mentions 'When too small an interval picked'; that part is the same defect as C018 and is not counted again here.
Second check: none (judge only)
Report a problem with this finding
Typo §7.3.1 misspells 'coarse-grained': "Tachyon only uses the coarse-gained recomputation at the job level".
- p. 11 §7.3.1 assumptions
coarse-gained recomputation
- p. 11
coarse-gained recomputation at the job level
Judge's note: Confirmed on the page 11 image; the same sentence uses 'fine-grained' correctly.
Second check: none (judge only)
Report a problem with this finding
Typo §7.3.2 uses a non-word: "recomputating lost data" (should be 'recomputing').
- p. 11 §7.3.2
recomputating lost data
Judge's note: Confirmed on the page 11 image, first line of §7.3.2.
Second check: none (judge only)
Report a problem with this finding
Typo §7.3.1 has a stray word: "Even if this data is recovered serially, and of all the other machines are blocked".
- p. 11 §7.3.1 Worst-case analysis
and of all the other machines are blocked
Judge's note: Confirmed on the page 11 image, Worst-case analysis paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo §7.2.2 uses the wrong word forms: "Networked bounded applications running in the same cluster" (should be 'Network-bound applications').
- p. 11 §7.2.2
Networked bounded applications
Judge's note: Confirmed on the page 11 image.
Second check: none (judge only)
Report a problem with this finding
Typo References [11] and [12] end with stray punctuation: "In USENIX HotOS 2011, ." and "In NSDI 2012, .".
- p. 13 References [11], [12]
In USENIX HotOS 2011, .
- p. 13 Reference [12]
In NSDI 2012, .
Judge's note: Visible in the typeset reference list on page 13 (crop), so it is not a text-extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [17] title is misspelled: "Lineage Retrieval for Scientic Data Processing" should read "Scientific".
- p. 14 Reference [17]
Lineage Retrieval for Scientic Data
Judge's note: The page image of p14 shows "Scientic", so this is not a lost 'fi' ligature in the text layer. The duplicate entry [18] spells the word "scientific" correctly.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [33] lists the author as "R. Rajikumar"; the name is probably "R. Rajkumar".
- p. 14 Reference [33]
R. Rajikumar
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the co-author's name is Ragunathan Rajkumar. The page image of p14 shows "R. Rajikumar" as quoted. Nothing in the paper allows an internal check.
Second check: none (judge only)