Periodic Reporting for period 3 - DUALITY (Theoretical Foundations of Memory Micro-Insertions in Wireless Communications)
Période du rapport: 2020-04-01 au 2021-09-30
What is the problem/issue being addressed?
------------------------------------------
We seek the theoretical foundations of turning “memory into data rates”, and to pursue the mathematical convergence between feedback information theory and preemptive distributed storage.
We seek to understand what is the best way to use caching in wireless communications, and to design fast memory-aided communication algorithms. We seek deep connections between feedback information theory and memory, and to identify the basic principles of how a splash of memory can surgically alter the informational structure of these networks, rendering them faster, simpler and more efficient.
Why is it important for society?
---------------------------------
Our project DUALITY seeks to reveal how, with a splash of feedback and smart caching of a mirco fraction of the popular content, we can handle the extreme increase in users and demand. DUALITY stands to transform the way we transmit data. It has the potential not only to benefit communication theory, but also network theory which currently has a somewhat divergent vision. The technological prospect of meaningfully turning memory into throughput, is exciting, and can translate Moore’s law for data storage into per-user throughput improvements, in some cases without requiring deployment of more infrastructure. This leads to an impactful turn of events: while more storage and users typically meant more traffic to worry about, now it can mean more cooperative memory to be translated into more throughput.
What are the overall objectives?
---------------------------------
The first objective of DUALITY is to explore the fundamental limits of memory-aided wireless communications. The key idea here is to derive the fundamental primitives that govern the use of preemptive distributed storage in multi-user wireless communications, paying special attention to understanding the intertwined relationship between feedback (quality-and-timeliness) and storage-capacity.
The second objective is to design algorithms that achieve memory-aided interference-management. The third scientific objective is to explore the basic principles of using memory to simplify implementation of communications. Another objective includes the study of the fundamental limits of memory-aided cooperation in large emerging networks such as the decentralized D2D networks, very large centralized networks (cloud-RAN or Massive MIMO), mm-Wave networks with ultra-high frequencies, and dense multi-cell networks with cell cooperation. The last objective involves wireless testbed experimentation of memory-aided communications algorithms.
------------------------------------------
We seek the theoretical foundations of turning “memory into data rates”, and to pursue the mathematical convergence between feedback information theory and preemptive distributed storage.
We seek to understand what is the best way to use caching in wireless communications, and to design fast memory-aided communication algorithms. We seek deep connections between feedback information theory and memory, and to identify the basic principles of how a splash of memory can surgically alter the informational structure of these networks, rendering them faster, simpler and more efficient.
Why is it important for society?
---------------------------------
Our project DUALITY seeks to reveal how, with a splash of feedback and smart caching of a mirco fraction of the popular content, we can handle the extreme increase in users and demand. DUALITY stands to transform the way we transmit data. It has the potential not only to benefit communication theory, but also network theory which currently has a somewhat divergent vision. The technological prospect of meaningfully turning memory into throughput, is exciting, and can translate Moore’s law for data storage into per-user throughput improvements, in some cases without requiring deployment of more infrastructure. This leads to an impactful turn of events: while more storage and users typically meant more traffic to worry about, now it can mean more cooperative memory to be translated into more throughput.
What are the overall objectives?
---------------------------------
The first objective of DUALITY is to explore the fundamental limits of memory-aided wireless communications. The key idea here is to derive the fundamental primitives that govern the use of preemptive distributed storage in multi-user wireless communications, paying special attention to understanding the intertwined relationship between feedback (quality-and-timeliness) and storage-capacity.
The second objective is to design algorithms that achieve memory-aided interference-management. The third scientific objective is to explore the basic principles of using memory to simplify implementation of communications. Another objective includes the study of the fundamental limits of memory-aided cooperation in large emerging networks such as the decentralized D2D networks, very large centralized networks (cloud-RAN or Massive MIMO), mm-Wave networks with ultra-high frequencies, and dense multi-cell networks with cell cooperation. The last objective involves wireless testbed experimentation of memory-aided communications algorithms.
Some of the work performed since the beginning of the project, can be found below.
Our main achievement so far is the combination of two opposing resources in wireless communications, namely multiple antennas and caching. The main outcome of our work is that pairing the multiplexing gains from multiple antennas with coded caching can be a very powerful combination, which can dramatically alleviate major fundamental limitations of both coded caching and multiple antenna precoding.
We managed to dramatically ameliorate the infamous subpacketization constraint of coded caching. Our work has shown that in practical scenarios, where the number of subpackets needs to be bounded, then L antennas can provide an L times higher speedup factor.
We resolved the bottleneck of having uneven cache sizes. In the extreme case where cache-aided users coexist with cache-less users, we have shown how a multi-antenna system can serve both user types at the same time, achieving optimal performance.
We resolved the CSIT bottleneck of cache-aided MIMO, by showing that the CSIT/CSIR costs can be untangled from the Coded Caching gains. In addition to the aforementioned savings, the proposed multi-antenna algorithm can further provide huge CSIT/CSIR savings by also reusing the already acquired feedback to transmit exponentially more information compared to the state-of-art.
We also resolved the uneven-channel bottleneck of wireless channels (where low-capacity users “slowing-down” the high-capacity ones) by identifying the information-theoretic fundamental limits of wireless Coded Caching with a single-antenna transmitter and users who experience different channel qualities.
Finally, we have extended our findings to distributed computing, and specifically to Coded MapReduce, a variant of the MapReduce distributed-computing framework, which uses increased job assignment at the nodes during the mapping phase in order to decrease the shuffling (communication) phase’s time.
DUALITY includes, in addition to different journal and conference publications, also two patent applications
- PCT patent application PCT/EP2018/073350 “System And Method For Managing Distribution Of Information In Multi-Antenna And Multi-Transmitter Environments”
- PCT patent application PCT/EP2018/083511. “System And Method For Managing Distribution Of Computations In Multi-Antenna And Multi-Transmitter Environments”
Our main achievement so far is the combination of two opposing resources in wireless communications, namely multiple antennas and caching. The main outcome of our work is that pairing the multiplexing gains from multiple antennas with coded caching can be a very powerful combination, which can dramatically alleviate major fundamental limitations of both coded caching and multiple antenna precoding.
We managed to dramatically ameliorate the infamous subpacketization constraint of coded caching. Our work has shown that in practical scenarios, where the number of subpackets needs to be bounded, then L antennas can provide an L times higher speedup factor.
We resolved the bottleneck of having uneven cache sizes. In the extreme case where cache-aided users coexist with cache-less users, we have shown how a multi-antenna system can serve both user types at the same time, achieving optimal performance.
We resolved the CSIT bottleneck of cache-aided MIMO, by showing that the CSIT/CSIR costs can be untangled from the Coded Caching gains. In addition to the aforementioned savings, the proposed multi-antenna algorithm can further provide huge CSIT/CSIR savings by also reusing the already acquired feedback to transmit exponentially more information compared to the state-of-art.
We also resolved the uneven-channel bottleneck of wireless channels (where low-capacity users “slowing-down” the high-capacity ones) by identifying the information-theoretic fundamental limits of wireless Coded Caching with a single-antenna transmitter and users who experience different channel qualities.
Finally, we have extended our findings to distributed computing, and specifically to Coded MapReduce, a variant of the MapReduce distributed-computing framework, which uses increased job assignment at the nodes during the mapping phase in order to decrease the shuffling (communication) phase’s time.
DUALITY includes, in addition to different journal and conference publications, also two patent applications
- PCT patent application PCT/EP2018/073350 “System And Method For Managing Distribution Of Information In Multi-Antenna And Multi-Transmitter Environments”
- PCT patent application PCT/EP2018/083511. “System And Method For Managing Distribution Of Computations In Multi-Antenna And Multi-Transmitter Environments”
The main progress beyond the SoA, draws from the main novelty behind our approach which has been to use caching not as a means of reducing the volume of the communication problem, but rather as a means for changing the structure of the problem.
We have addressed the existing crippling bottlenecks of cache-aided communications.
One considerable improvement over the SoA can be found in our effort to alleviate the subpacketization bottleneck which can severely deteriorate the Coded Caching gains in practical systems. The progress here was that -- as we have shown for the first time -- pairing transmitters with Coded Caching not only does not increase the required subpacketization, but actually can severely reduce it. Our contribution beyond SoA lies in the realization that having this extra dimensionality on the transmitter side, in fact reduces rather than increases subpacketization, and does so in a very accelerated manner. This property is based on the novel principle of the virtual decomposition of the cache-aided MISO BC into parallel, single-stream coded caching channels with fewer users each. This decomposition is made possible because, as we have shown, the optimal performance can be gained without encoding across parallel channels.
The second progress beyond SoA involved our effort to alleviate the CSI Bottleneck, which saw a scaling (with the number of users) of the feedback requirements of multi-antenna systems. The effects of this bottleneck were first revealed in our work, which provided the novelty of a fundamentally different XOR design, which managed to reduce the per-transmission CSI requirements, from CSI that scaled with the number of users, to CSI equal to L, i.e. managed to completely untangled the CSI cost from the number of users.
Similar progress involved completely new ways of designing algorithms to resolve the link-asymmetry bottleneck. It was long thought that the performance of coded caching is limited by the weakest-link user. We provided a completely novel solution that bypasses information theoretic intuition by sending sequences of multi-casting messages, one on top of each other, thus showing that in fact the weakest users need not “pull down” the rest, but rather that the strongest users can “pull up” the rest, without any performance degradation for themselves.
We have addressed the existing crippling bottlenecks of cache-aided communications.
One considerable improvement over the SoA can be found in our effort to alleviate the subpacketization bottleneck which can severely deteriorate the Coded Caching gains in practical systems. The progress here was that -- as we have shown for the first time -- pairing transmitters with Coded Caching not only does not increase the required subpacketization, but actually can severely reduce it. Our contribution beyond SoA lies in the realization that having this extra dimensionality on the transmitter side, in fact reduces rather than increases subpacketization, and does so in a very accelerated manner. This property is based on the novel principle of the virtual decomposition of the cache-aided MISO BC into parallel, single-stream coded caching channels with fewer users each. This decomposition is made possible because, as we have shown, the optimal performance can be gained without encoding across parallel channels.
The second progress beyond SoA involved our effort to alleviate the CSI Bottleneck, which saw a scaling (with the number of users) of the feedback requirements of multi-antenna systems. The effects of this bottleneck were first revealed in our work, which provided the novelty of a fundamentally different XOR design, which managed to reduce the per-transmission CSI requirements, from CSI that scaled with the number of users, to CSI equal to L, i.e. managed to completely untangled the CSI cost from the number of users.
Similar progress involved completely new ways of designing algorithms to resolve the link-asymmetry bottleneck. It was long thought that the performance of coded caching is limited by the weakest-link user. We provided a completely novel solution that bypasses information theoretic intuition by sending sequences of multi-casting messages, one on top of each other, thus showing that in fact the weakest users need not “pull down” the rest, but rather that the strongest users can “pull up” the rest, without any performance degradation for themselves.