A Scalable, Commodity Data Center Network Architecture
ACM SIGCOMM Computer Communication Review, 2008 · paper · Computer science
Shows how to build data center networks from mostly commodity Ethernet switches that support full aggregate bandwidth for tens of thousands of machines. The authors argue this can deliver more performance at lower cost, while staying compatible with Ethernet, IP and TCP.
Checked: PDF from cseweb.ucsd.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
4 findings
Substance Routing example: "forwards on port 3, connected to core switch 10.4.1.1", but Figure 3's dashed path runs from 10.0.2.1 to 10.4.1.2, and Figure 5 maps port 3 to 10.4.1.2.
- p. 6 Section 3.5, Routing Example
forwards on port 3, connected to core switch 10.4.1.1
- p. 5 Figure 3 (dashed path) and Figure 5 RAM table
10.4.1.2
Judge's note: In the Figure 3 image, the dashed line goes from 10.0.2.1 to 10.4.1.2 and then to 10.2.2.1. The Figure 5 RAM table maps output port 3 to next hop 10.4.1.2, and port 2 to 10.4.1.1. The text on page 6 says 10.4.1.1.
Second check: another model agreed
Report a problem with this finding
Substance Algorithm 2 line 3, "foreach destination pod x in [0, (k/2) − 1] do", covers only half of the k pods; it should be [0, k − 1].
- p. 6 Algorithm 2, line 3
foreach destination pod x in [0, (k/2) −1] do
- p. 6 Core Switches paragraph
Since each core switch is connected to every pod (port i is con-nected to pod i)
Judge's note: Page 6, checked in the image. Section 3.2 numbers pods over [0, k−1]. The text says each core switch connects to every pod ("port i is connected to pod i") and holds up to k first-level prefixes. The routing example needs 10.2.0.0/16 for k=4, which the printed loop range (pods 0 and 1) never generates.
Second check: another model agreed
Report a problem with this finding
Substance Section 3.9 states the fat-tree has "56.6% less power consumption and 56.5% less heat dissipation", but the Figure 7 bars (about 728 vs about 248 kW; about 3160 vs about 1080 kBTU/hr) both indicate a reduction of about 66%.
- p. 8 Section 3.9
56.6% less power consumption and 56.5% less heat dissipation
- p. 8 Figure 7
Hierarchical design
Only the authors can say what is right here.
Judge's note: Viewed the Figure 7 crop at high resolution. Hierarchical total power is about 728 kW and fat-tree about 248 kW. Heat bars on the left scale are about 316 and 108, which are about 3160 and 1080 kBTU/hr on the right axis. Both reductions come to about 66%, roughly 9 points above the printed 56.6%/56.5%. Reading error on the bars is only a few percent and cannot close that gap. The printed figures may be transposed digits (65.x%), but only the authors can say which value is correct.
Second check: another model agreed
Report a problem with this finding
Reference or label Wrong section pointer: "we have built a prototype of our architecture in Click [21] as described in Section 3". The Click prototype is described in Section 4 (Implementation).
- p. 4 Section 2.3 Summary
built a prototype of our architecture in Click [21] as described in Section 3.
- p. 4 Section 2.3 Summary, last sentences
built a prototype of our architecture in Click [21] as described in Section 3
- p. 8 Section 4 Implementation
To carry out larger-scale evaluations, we also built a prototype using Click
- p. 4 Section 2.3
Click [21] as described in Section 3.
Judge's note: Page 4, Section 2.3, checked in the image. Section 3 is "Architecture". The text of Section 4 "Implementation" (page 8) describes the Click prototype ("we also built a prototype using Click", the TwoLevelTable element).
Second check: another model agreed
Report a problem with this finding
18 minor (typos and the like; most checked by the judge only)
Wording The definition of oversubscription, "the ratio of the worst-case achievable aggregate bandwidth among the end hosts to the total bisection bandwidth", is inverted relative to how it is used ("5:1 means that only 20% of available host bandwidth is available").
- p. 2 Section 2.1.2
ratio of the worst-case achievable aggregate bandwidth among the end hosts to the total bisection bandwidth
Judge's note: Section 2.1.2, page 2, checked in the image. Read literally, the ratio achievable:total gives 1:5 for 20% availability, but the paper writes 5:1, and 2.5:1 for 400 Mbps. The order of the two terms in the definition is swapped. The intended meaning is clear from context.
Second check: another model agreed
Report a problem with this finding
Wording "would be $8.64M, compared to $37M for the traditional techniques described earlier": the $8.64M is for 27,648 hosts, while the $37M (Section 2.1.4) is for about 20,000 hosts, and the text does not mention the difference in cluster size.
- p. 4 Section 2.2, end of fat-tree example
would be $8.64M, compared to $37M for the tradi- tional techniques described earlier
- p. 2 Section 2.1.4
interconnect 20,000 hosts with full bandwidth among all hosts comes to approximately $37M
Judge's note: Page 4, Section 2.2, and page 2, Section 2.1.4 ("interconnect 20,000 hosts ... approximately $37M"). Table 1 for 2008 gives 20,480 hosts at $1.8K per port and 27,648 hosts for the fat-tree. Both figures are the largest configuration of each design, so the comparison is defensible. It is not like-for-like, though, and the sentence does not say so.
Second check: another model agreed
Report a problem with this finding
Wording "the routing table of any pod switch will contain no more than k/2 prefixes and k/2 suffixes", yet Algorithm 1 adds k/2 /24 prefixes plus 0.0.0.0/0, and Figure 4 (k=4) shows 3 prefixes.
- p. 5 Section 3.3, last sentence
than k/2 prefixes and k/2 suffixes.
- p. 5 Figure 4
0.0.0.0/0
- p. 5 Section 3.3, last sentence
than k/2 prefixes and k/2 suffixes
Only the authors can say what is right here.
Judge's note: Page 5, Figure 4, checked in the image: it lists 10.2.0.0/24, 10.2.1.0/24 and 0.0.0.0/0, i.e. 3 prefixes where k/2 = 2. Algorithm 1 (page 6) lines 3–6 add k/2 + 1 first-level prefixes. Unless the default /0 entry is not meant to count, the bound is off by one.
Second check: another model agreed
Report a problem with this finding
Wording Section 3.9 says "Figure 6 plots the average over three different switch models", but Figure 6 shows separate bars for six named switch models (Catalyst 4948, ProCurve 2900-48G, Netgear GSM 7352S, Catalyst 6509-E, BigIron RX-32, BlackDiamond 10808) and plots no averages.
- p. 8 Section 3.9 and Figure 6
Figure 6 plots the average over three different switch mod-els.
- p. 8 Section 3.9
Figure 6 plots the average over three different switch mod- els
- p. 8 Figure 6 x-axis
Catalyst 4948
Only the authors can say what is right here.
Judge's note: Viewed page 8. The Figure 6 x-axis has six individual switch labels, each with a Watts/Gbps bar and a BTU/hr/Gbps bar. The next sentence refers to "the last three on the x-axis", so the plot shows individual models. The authors probably mean that the comparison averages three models per class, but as written the sentence misdescribes what the figure shows.
Second check: another model agreed
Report a problem with this finding
Wording Section 4.3 describes the scheduler's array as "of all the links in the network (a total of 4 ∗ k ∗ (k/2)^2 links)"; a k-ary fat-tree has 3k(k/2)^2 undirected links including host links, so the count fits only an unstated convention (e.g. directed switch-to-switch links).
- p. 9 Section 4.3
a total of 4 ∗k ∗(k/2)2 links
Only the authors can say what is right here.
Judge's note: Viewed the page 9 crop. Counting undirected links: k^3/4 host-edge, k^3/4 edge-aggregation and k^3/4 aggregation-core, so 3k^3/4 in total. The printed 4k(k/2)^2 = k^3 equals the number of directed inter-switch links (2 x 2 x k^3/4), which is a plausible meaning for a reservation array. It is consistent with Table 3 (64 B at k=4, 111 KB at k=48 at one byte per entry). The paper does not say which links are counted or whether direction matters, so "all the links in the network" is ambiguous rather than clearly wrong.
Second check: another model agreed
Report a problem with this finding
Wording Section 5.1 reads "four machines each running four pod switches with one additional uplink. The four pod switches are connected to a 4-port core switch"; read literally, that is 16 pod switches, but the next sentence says there are four.
- p. 9 Section 5.1, second paragraph
each running four pod switches with one additional uplink. The four pod switches are connected to a 4-port core switch
Only the authors can say what is right here.
Judge's note: Viewed page 9, right column. With 16 hosts and a 4-port core, the context implies one pod switch per machine (likely a four-port pod switch plus an uplink), but the sentence as printed is self-contradictory.
Second check: another model agreed
Report a problem with this finding
Wording Section 5.3 says the FlowScheduler achieves "outperforming all other methods in all the benchmark tests", but Table 2 shows Flow Scheduling at 99.5% (Stride 2) and 99.9% (Stride 8), below the 100.0% of Two-Level Table and Flow Classification, and only tied at 100.0% on Stride(1), Stride(4) and Staggered (1.0, 0.0).
- p. 10 Section 5.3, paragraph on FlowScheduler
outperforming all other methods in all the bench-mark tests
- p. 10 Table 2, Stride (2) and Stride (8) rows
99.5%
- p. 10 Section 5.3 Results
outperforming all other methods in all the bench-
Judge's note: Viewed Table 2 on page 10. The values are as described. The claim of outperforming in every test overstates the table, though the differences are small.
Second check: another model agreed
Report a problem with this finding
Wording Section 5.3 says "the flow classifier outperforms both the traditional tree and the two-level table in all cases", but in Table 2 it ties with Two-Level Table at 100.0% on Stride (1), (2), (4), (8) and Staggered (1.0, 0.0), and ties with Tree on Stride (1) and Staggered (1.0, 0.0).
- p. 10 Section 5.3
level table in all cases, with a worst-case bisection bandwidth of
- p. 10 Table 2, Stride rows
100.0%
- p. 10 Section 5.3, paragraph on flow classifier
level table in all cases, with a worst-case bisection bandwidth of approximately 75%
Judge's note: Viewed Table 2 on page 10. Flow Classification is never worse than either baseline but ties in five rows, so "outperforms ... in all cases" overstates the result. The quoted worst case of about 75% (75.1%, 75.4%) does match the table.
Second check: another model agreed
Report a problem with this finding
Typo Wrong word: "Today, the principle bottleneck in large-scale clusters" should read "principal bottleneck".
- p. 1 Section 1, second paragraph
Today, the principle bottleneck in large-scale clusters
- p. 1 Introduction, para 2
the principle bottleneck in large-scale clusters
Judge's note: Page 1, right column, Introduction paragraph 2: the image shows "principle bottleneck".
Second check: none (judge only)
Report a problem with this finding
Typo Subject–verb agreement: "Growing expertise with clusters of commodity PCs have enabled" should read "has enabled".
- p. 1 Section 1, first sentence
Growing expertise with clusters of commodity PCs have enabled
Judge's note: Page 1, first sentence of the Introduction, checked in the image. The subject "expertise" is singular.
Second check: none (judge only)
Report a problem with this finding
Typo "Important applications classes include" should read "application classes".
- p. 1 Introduction
Important applica- tions classes
Judge's note: Page 1, right column, top: the image shows "applica-tions classes".
Second check: none (judge only)
Report a problem with this finding
Typo Missing word: "capable of providing full bandwidth to up 27,648 hosts" should read "up to 27,648 hosts".
- p. 2 Section 1, after goals list
full bandwidth to up 27,648 hosts
Judge's note: Page 2, left column, checked in the image. The word is missing in print.
Second check: none (judge only)
Report a problem with this finding
Typo "Tables 1 also shows the largest commodity GigE switch" should read "Table 1 also shows".
- p. 3 Section 2.1.4, last paragraph
Tables 1 also shows
- p. 3 Section 2.1.4
Tables 1 also shows the largest commodity GigE switch
Judge's note: Page 3, left column, bottom: the plural is visible in the image.
Second check: none (judge only)
Report a problem with this finding
Typo Subject–verb agreement: "the core switches contains only terminating /16 prefixes" should read "contain".
- p. 6 Core Switches paragraph
the core switches contains only terminating /16
Judge's note: Page 6, Core Switches paragraph, checked in the image.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.8 reads "(b) between core and a upper-level switches"; the stray article "a" is ungrammatical (should be "between core and upper-level switches").
- p. 7 Section 3.8, second paragraph
between core and a upper-level switches
Judge's note: Checked the crop of page 7, right column, Section 3.8. The text reads exactly as quoted. This is a copy-editing slip.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.6 reads "step 2 aims to ensure fair distribution on flows on upward-pointing ports"; "distribution on flows" should read "distribution of flows".
- p. 7 Section 3.6 after list
fair distribution on flows on upward-pointing ports
Judge's note: Checked the crop of page 7, left column. The wording is as quoted. It is a minor preposition error.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.9 writes "the cheaper NetGear switch" one sentence after "2,880 Netgear GSM 7252S switches"; the vendor name is capitalised inconsistently.
- p. 8 Section 3.9
We are able to use the cheaper NetGear switch
Judge's note: Checked the crop of page 8, right column. Both spellings appear in adjacent sentences, and Figure 6 uses "Netgear". This is cosmetic only.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [15] reads "Tree Bitmap : Hardware/Software IP Lookups with Incremental Updates." with a stray space before the colon.
- p. 12 Reference [15]
Tree Bitmap :
Judge's note: Checked the crop of the page 12 reference list. The space before the colon is visible. This is cosmetic only.
Second check: none (judge only)