SCDP: Systematic Rateless Coding for Efficient Data Transport in Data Centres (Complete Version)
read more
Citations
Reducing tail latency with coding-based packet spraying in edge datacenters
A Cost-Effective and Multi-Source-Aware Replica Migration Approach for Geo-Distributed Data Centers
References
MapReduce: simplified data processing on large clusters
The Google file system
The part-time parliament
BCube: a high performance, server-centric network architecture for modular data centers
Ceph: a scalable, high-performance distributed file system
Related Papers (5)
One packet suffices - Highly efficient packetized Network Coding With finite memory
Reliable data delivery over deep space networks: Benefits of long erasure codes over ARQ strategies
Frequently Asked Questions (10)
Q2. What future works have the authors mentioned in the paper "Scdp: systematic rateless coding for efficient data transport in data centres" ?
As part of their future work, the authors aim at developing an SCDP prototype ( in-kernel and/or using user-space network stack ) and exploring its performance with real application workloads. As part of their future work, the authors will investigate this argument further by developing extensions of existing unicast data centre protocols ( e. g. [ 26 ] ) that can handle one-to-many and many-to-one data transport and compare their performance with SCDP.
Q3. How many aggregation switches are in each pod?
For their experimentation the authors have used a 250-server FatTree topology with 25 core switches and 5 aggregation switches in each pod (50 aggregation switches in total).
Q4. What is the reason why PIAS performs worse than NDP?
In general, the authors argue that PIAS performs worse than NDP because (1) it relies on DCTCP for data transport and as a result it suffers from the limitations of a single-path protocol (i.e. lack of support for multi-path transport and packet spraying); (2) connection establishment requires a three-way handshake and senders start with a small window, both of which can severely hurt FCTs for short flows; and (3) buffer occupancy in NDP is significantly lower than in PIAS [26] which also affects performance for short slows.
Q5. Why is the gap between SCDP and NDP at its smallest?
When the network load is very high, the gap between SCDP and NDP is at its smallest, because losses (trimmed packets) and therefore decoding are more frequent.
Q6. How do the authors model the decoding latency?
The authors model the decoding latency based on the results reported in [55], by fitting the worst-case decoding latencies for different number of K source symbols into a polynomial function.
Q7. How does SCDP reduce the decoding overhead for large data blocks?
This is done by employing pipelining of source blocks, which alleviates the decoding overhead for large data blocks and maximises application goodput (see Section IV-F).
Q8. How will the authors investigate this argument further?
As part of their future work, the authors will investigate this argument further by developing extensions of existing unicast data centre protocols (e.g. [26]) that can handle one-to-many and many-to-one data transport and compare their performance with SCDP.
Q9. What does SCDP do to reduce the latency of the source block?
This ensures that SCDP does not induce any unnecessary overhead; i.e. symbol packets that are redundant in decoding the source block.
Q10. What is the difference between NDP and SCDP?
For higher loads, NDP performs even worse than SCDP because of the lack of support for MLFQ, which results in the trimming of more packets belonging to short flows.