TCP incast solutions in data center networks: A classification and survey

Alipio, Melchizedek I.; Tiglao, Nestor Michael C.; Bokhari, Fawaz; Khalid, Salman

doi:10.1016/j.jnca.2019.102421

Cited by 18 publications

(14 citation statements)

References 50 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For large message sizes from a few workers, 80.11% of the maximum possible goodput can be achieved. But as the number of workers increases, we observe the onset of incast [2,5], whereby goodput collapses because of recurring congestion and TCP timeouts due to contention between the large number of senders. Especially for large number of workers, this problem is not solvable using standard TCP congestion window adjustments: for 1000 workers, a switch buffer of 1 MB will be overwhelmed with even just 1 packet per worker, thus requiring the equivalent of a congestion window less than 1 packet.…”

Section: Motivationmentioning

confidence: 91%

“…TCP incast is a well-studied problem in the TCP-on-CPU setting [2,5]. Readers familiar with such work may even suspect that offloading scatter-gather to the FPGA would worsen incast: the FPGA can send out queries to workers faster, and thus perhaps elicit responses in a near-synchronous fashion, thus accelerating incast onset, and making offload not only futile, but actively harmful.…”

Section: Simple and Effective Schedulingmentioning

confidence: 99%

“…Incast: Incast is a well-recognized problem in data center workloads with high fan-out. We refer interested readers to comprehensive surveys detailing the problem's history and the numerous approaches proposed to tackle it [2,5].…”

Section: Related Workmentioning

confidence: 99%

“…This time scales linearly with the number of workers, such that for 1000 workers, it would exceed 10 ms. Further, fetching only tens of kilobytes of data from each of only 100 workers saturates 2 cores on a 4-core server ( §2). The gather operation also suffers from the well-understood incast phenomenon [2,5], whereby a large number of workers sending data to a receiver congest network ports towards the receiver, resulting in high packet loss and TCP timeouts. This can slow down the communication by orders of magnitude compared to the line-rate.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Specializing the network for scatter-gather workloads

Álvarez

Alonso

et al. 2020

Proceedings of the 11th ACM Symposium on Cloud Computing

View full text Add to dashboard Cite

Data processing and distributed querying workloads often involve a "scatter-gather" or "partition-aggregate" architectural pattern, whereby one application queries hundreds or even thousands of workers. Network communication is often a bottleneck in this pattern, especially when the compute task at each worker is small, such as for Web queries and interactive analytics. The network bottleneck can result in low throughput, high CPU utilization, and cause job completion time to increase by orders of magnitude. To overcome these inefficiencies, we explore hardware-offload of the scatter-gather primitive, whereby a smart NIC takes on the responsibility of sending out queries and collecting responses. We show that this approach not only virtually eliminates CPU usage, but with suitable scheduling of responses, it also speeds up scatter by allowing parallel queries, and gather by preventing throughput collapse due to excessive congestion. Besides response scheduling, we use a careful design at the NIC to limit FPGA resource usage: our approach uses about 25% of on-chip logic and 33% of on-chip memory on a midsized FPGA, leaving enough room for implementing other functions on the smart NIC.

show abstract

Section: Motivationmentioning

confidence: 91%

Section: Simple and Effective Schedulingmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Specializing the network for scatter-gather workloads

Álvarez

Alonso

et al. 2020

Proceedings of the 11th ACM Symposium on Cloud Computing

View full text Add to dashboard Cite

show abstract

“…Many data centers are used to support the cloud computing, which constructs a Data Center Network (DCN). In other words, DCN is the main backbone of cloud computing 2 . Nowadays, with the constant emerging of bandwidth‐hungry applications and networking paradigms (eg, mobile Internet and IoT), the volume of traffic to be handled in DCN has shown the explosive increasing tendency.…”

Section: Introductionmentioning

confidence: 99%

Tennis‐resource allocation mechanism in software‐defined data center network

Zhang

2021

Internet Technology Letters

View full text Add to dashboard Cite

With the rapid development of cloud computing, most network resources are usually stored at Data Center Network (DCN) for providing the remote storage services. Under such paradigm, there arises some issues, such as traffic scheduling, energy saving and resource allocation. At present, the resource allocation in DCN has attracted the extensive concern. In order to effectively manage and allocate these resources in DCN, Software‐Defined Networking can be employed to achieve such purpose, by leveraging its natural abilities for the global control and centralized management. Another point, the tennis is a very popular sport to enhance the physical fitness. However, the tennis‐resources are considerably limited as the building of venue is quite expensive. Therefore, it becomes very significant to investigate the tennis‐resource allocation in Software‐Defined DCN (SDCN). Consider that there are many tenants using the tennis‐resources in SDCN, this paper studies the multi‐tenant tennis‐resource allocation scenario. Meanwhile, the Global Superiority Equalization algorithm is devised. At first, the max‐min fairness is used to determine the allocated number of tennis‐resources for each round. Then, the allocation state is adjusted according to the current situation until the final allocation decision is determined. The experiment results show that the proposed tennis‐resource allocation mechanism can reduce the resource allocation cost and improve the resource utilization. Accordingly, SDCN can handle more tennis‐resource requests and increase the benefit of venue provider.

show abstract