Tesi etd-10042025-105110
Link copiato negli appunti
Tipo di tesi
Corso Ordinario Secondo Livello
Autore
CARLON, JACOPO
URN
etd-10042025-105110
Titolo
Energy-efficient low-latency packet processing exploiting OS Kernel Bypass with DPDK
Struttura
Classe Scienze Sperimentali
Corso di studi
INGEGNERIA - INGEGNERIA
Relatori
tutor Prof. FORESTIERI, ENRICO
relatore Prof. CUCINOTTA, TOMMASO
relatore Prof. LETTIERI, GIUSEPPE
relatore Prof. CUCINOTTA, TOMMASO
relatore Prof. LETTIERI, GIUSEPPE
Parole chiave
- DPDK
- Kernel Bypass
- Packet Forwarding
- Power- Awareness
- Ultra-LowLatency
Data inizio appello
11/12/2025;
Disponibilità
parziale
Riassunto analitico
In the last decades, networking services have evolved to impose low latency as a standard requirement.
With the recent trend towards abstracting the network architecture by softwarizing its services, it has become common to use a Linux operating system for their implementation. However, its kernel's inherent comprehensiveness and generality in packet processing cause it to fail at providing the best possible latency performance.
Over the years, research has evolved culminating in a number of kernel bypass techniques.
In this thesis, the Data Plane Development Kit (DPDK) has been leveraged in order to study the tradeoff between low-latency L3 packet processing and its associated power consumption.
In this work, a number of alternatives have been designed, implemented, evaluated and tested against realistic traffic patterns.
The best solution developed was obtained by abandoning the costly busy-polling implemented by DPDK's Poll Mode Driver (PMD), and by instead alternating between an efficient polling-pausing phase with a far more energy-saving phase during which interrupts are enabled (which is exited with a dynamically adjusted timeout, in order to avoid latency spikes).
This novel implementation has greatly narrowed the gap between the most power-aware and the most latency-performing existing approaches.
With the recent trend towards abstracting the network architecture by softwarizing its services, it has become common to use a Linux operating system for their implementation. However, its kernel's inherent comprehensiveness and generality in packet processing cause it to fail at providing the best possible latency performance.
Over the years, research has evolved culminating in a number of kernel bypass techniques.
In this thesis, the Data Plane Development Kit (DPDK) has been leveraged in order to study the tradeoff between low-latency L3 packet processing and its associated power consumption.
In this work, a number of alternatives have been designed, implemented, evaluated and tested against realistic traffic patterns.
The best solution developed was obtained by abandoning the costly busy-polling implemented by DPDK's Poll Mode Driver (PMD), and by instead alternating between an efficient polling-pausing phase with a far more energy-saving phase during which interrupts are enabled (which is exited with a dynamically adjusted timeout, in order to avoid latency spikes).
This novel implementation has greatly narrowed the gap between the most power-aware and the most latency-performing existing approaches.
File
| Nome file | Dimensione |
|---|---|
Ci sono 1 file riservati su richiesta dell'autore. |
|