F2FS: A New File System for Flash Storage
File and Storage Technologies, 2015 · paper · Computer science
Describes F2FS, a Linux file system built on append-only logging and designed around flash storage characteristics. The authors report beating EXT4 by up to 3.1× on a mobile system and up to 2.5× on a server SATA SSD.
Checked: PDF from usenix.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Substance Text says NILFS2's periodic flushes "limited the performance gain over EXT4 to 1.8×" on the SATA SSD fileserver workload, but the NILFS2 fileserver bar in Figure 4(a) is at about 1.5.
- p. 9 Section 3.2.2, fileserver paragraph
limited the performance gain over EXT4 to 1.8×
- p. 9 Figure 4(a) SATA SSD, fileserver NILFS2 bar
(a) SATA SSD
Only the authors can say what is right here.
Judge's note: Checked Figure 4(a) on p9. The NILFS2 (hatched) fileserver bar tops out at the 1.5 gridline; its error bar reaches about 1.8. The 1.8 in the text therefore matches the top of the error bar, not the plotted bar. The sentence follows the Figure 5(d) SATA trace discussion, so it refers to the SATA SSD. On the PCIe SSD (Fig. 4(b)), the NILFS2 fileserver bar is about 0.63. Authors should say which value is meant.
Second check: another model agreed
Report a problem with this finding
Substance "BTRFS and F2FS (especially F2FS_normal) saw a severe performance drop of 22% and 48% each" and "22% (comparable to BTRFS...)", but in Figure 8(a) BTRFS falls from 1.0 to about 0.52 (about 48%) and F2FS_normal from about 2.17 to about 1.28 (about 41%).
- p. 12 Section 3.2.5, fileserver test
saw a se-vere performance drop of 22% and 48% each
- p. 12 Figure 8(a)
(a) fileserver (random operations) on a device filled up 94%.
- p. 12 Figure 8(a)
fileserver (random operations) on a device filled up 94%.
- p. 12 3.2.5 fileserver test
(comparable to BTRFS and not too far from EXT4)
Only the authors can say what is right here.
Judge's note: Read Figure 8(a) on p12 by eye. EXT4 goes 1.0→0.83 (17%, matching the text). F2FS_adaptive goes ~2.17→~1.72 (~21%, matching 22%). BTRFS goes 1.0→~0.52 (~48%), not 22%. F2FS_normal goes ~2.17→~1.28 (~41%). The 22% given for BTRFS, and the 'comparable to BTRFS' comparison that rests on it, do not match the plotted BTRFS series. The percentages may have been swapped or misassigned; only the authors can say.
Second check: another model agreed
Report a problem with this finding
7 minor (typos and the like; most checked by the judge only)
Wording oltp thread count: Section 3.1 says it "updates their data randomly with fsync with 200 threads in parallel", but the Table 3 oltp row lists 211 threads.
- p. 8 Section 3.1 text
randomly with fsync with 200 threads in parallel
- p. 8 Table 3, oltp row, Threads column
211
- p. 8 3.1 text
with 200 threads in parallel
Only the authors can say what is right here.
Judge's note: On page image p8, the Table 3 oltp row reads Threads = 211, and the text below says 200 threads. The 200 may count only the update threads, with 211 including helper (writer) threads, but the paper does not say so. Asking the authors to reconcile the two counts.
Second check: another model agreed
Report a problem with this finding
Wording Section 3.2.2 says oltp writes "on a single 800MB database file", while Section 3.1 says "oltp pre-allocates ten large files" and Table 3 lists Files = 10 for oltp.
- p. 10 Section 3.2.2, oltp paragraph
on a single 800MB database file
- p. 8 Section 3.1 text
oltp pre-allocates ten large files
- p. 8 Table 3, oltp row, Files column
10
- p. 10 Section 3.2.2, oltp paragraph
fsync calls on a single 800MB database file
Only the authors can say what is right here.
Judge's note: Checked p8: the Table 3 oltp row reads Files = 10, File size = 800MB, and the text says 'ten large files'. Checked p10: 'fsync calls on a single 800MB database file'. The two descriptions of the same workload are not reconciled anywhere in the paper.
Second check: another model agreed
Report a problem with this finding
Wording "On the PCIe SSD, all file systems perform rather similarly" for fileserver, but in Figure 4(b) BTRFS is at about 0.78 and NILFS2 at about 0.63, against EXT4 = 1.0 and F2FS ≈ 1.03.
- p. 9 Section 3.2.2, end of fileserver paragraph
PCIe SSD, all file systems perform rather similarly
- p. 9 Figure 4(b) PCIe SSD, fileserver bars
(b) PCIe SSD
Only the authors can say what is right here.
Judge's note: Checked Figure 4(b) on p9. The F2FS and EXT4 fileserver bars are about 1.0, the BTRFS bar is about 0.78 and the NILFS2 bar about 0.63. Two error bars float above them at about 1.3 and 1.75, apparently detached from these bars, which is itself unclear. The wording 'all file systems ... rather similarly' overstates what the bars show; it may be meant for F2FS against EXT4 only.
Second check: another model agreed
Report a problem with this finding
Typo Subject–verb agreement: "Inlining reduces space requirements and improve I/O performance" should read "improves".
- p. 4 Section 2.2, second column
and improve I/O performance
Judge's note: Checked on the p4 crop: the printed text says 'and improve I/O performance' after the singular subject 'Inlining'.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 7 (upper panel) y-axis label reads "Normalized Performace"; it should read "Performance".
- p. 11 Figure 7 upper panel, y-axis
Normalized Performace
Judge's note: Checked the p11 crop: the misspelling is printed in the figure's axis label.
Second check: none (judge only)
Report a problem with this finding
Typo Mid-sentence capital: "Still, We measured over 10% performance improvement" should read "we".
- p. 12 Section 3.2.4, last paragraph
Still, We measured over
Judge's note: Seen on page image p12, Section 3.2.4, last paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo References [8] and [11] read "USENIX Anual Technical Conference (ATC)"; it should read "Annual".
- p. 14 Reference [8]
USENIX Anual Technical Conference (ATC)
- p. 14 Reference [11]
of the USENIX Anual Technical Conference (ATC)
- p. 14 References [11]
of the USENIX Anual Technical Conference (ATC),
Judge's note: Seen on p14 for [8] (crop) and confirmed in the text layer for both [8] and [11].
Second check: none (judge only)