TY - GEN
T1 - CacheNet
T2 - 20th Annual IFIP Networking Conference, IFIP Networking 2021
AU - Griner, Chen
AU - Avin, Chen
N1 - Funding Information:
Figure 1 (b) shows the results for the Hadoop trace. They present a case were CacheNet was able to improve on rotor-net, but not nearly as significantly as for the DB trace. In particular we see that LFU and LRU were able to reach about 10% improvement with a small cache of about 5 to 10 switches, while the improvement brought by OPT is more significant at around 20%. These results with Hadoop are surprising, since the Hadoop trace lacks significant structure [6], which should lead to negligible improvement. However, looking at the hit ratio curves where x > 16 the hit ratio seems to grow at a slightly faster rate, which may indicate some structure that LFU and OPT are able to use. Figure 1 (c) shows an HPC trace of the MultiGrid application [7]. The effectiveness plot shows that the hit ratio of LRU is superior to LFU. However, all three algorithms reach a hit rate of about 100% with 20 switches. One possible explanation for the under-performance of LFU is that while the HPC trace distributions are skewed, they are only skewed in the sense that they are sparse; that is, only a small part of the possible communicating pairs appear in the trace. The pairs that do appear in the trace are (relatively) uniformly distributed. Acknowledgement: We would like to thank Stefan Schimd for many discussions and his insightful feedback. This project received funding by the European Research Council (ERC), grant agreement no. 864228, Horizon 2020, 2020-2025.
Publisher Copyright:
© 2021 IFIP.
PY - 2021/6/21
Y1 - 2021/6/21
N2 - Emerging optical communication technologies support the dynamic reconfiguration of datacenter network topologies depending on the traffic they serve. However, to reap the benefits of such demand-aware networks, a control logic is required which allows to quickly learn and adapt to traffic patterns. This paper presents CacheNet, a novel approach to efficiently control demand-aware networks. CacheNet leverages temporal and spatial locality in the traffic by managing the reconfigurable links of the optical switches as a links-cache. Network traffic, in turn, can be served either by a link from the link-cache component or by a demand-oblivious topology component. We study several classic caching algorithms and provide an analytical model which captures their performance benefits compared to an all demand-oblivious topology. Our analytical results show that based on the hit ratios and the links-cache size, our hybrid design can outperform designs that are based only on demand-oblivious topology.
AB - Emerging optical communication technologies support the dynamic reconfiguration of datacenter network topologies depending on the traffic they serve. However, to reap the benefits of such demand-aware networks, a control logic is required which allows to quickly learn and adapt to traffic patterns. This paper presents CacheNet, a novel approach to efficiently control demand-aware networks. CacheNet leverages temporal and spatial locality in the traffic by managing the reconfigurable links of the optical switches as a links-cache. Network traffic, in turn, can be served either by a link from the link-cache component or by a demand-oblivious topology component. We study several classic caching algorithms and provide an analytical model which captures their performance benefits compared to an all demand-oblivious topology. Our analytical results show that based on the hit ratios and the links-cache size, our hybrid design can outperform designs that are based only on demand-oblivious topology.
UR - http://www.scopus.com/inward/record.url?scp=85112804910&partnerID=8YFLogxK
U2 - 10.23919/IFIPNetworking52078.2021.9472808
DO - 10.23919/IFIPNetworking52078.2021.9472808
M3 - Conference contribution
AN - SCOPUS:85112804910
T3 - 2021 IFIP Networking Conference, IFIP Networking 2021
BT - 2021 IFIP Networking Conference, IFIP Networking 2021
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 21 June 2021 through 24 June 2021
ER -