Your SlideShare is downloading.
×

- 1. Distributed Systems in the Post-Moore Era Dr. Vincenzo De Maio vincenzo@ec.tuwien.ac.at FWF START Prize 2015 http://rucon.ec.tuwien.ac.at/ TEWI KOLLOQUIUM Klagenfurt, 14th March 2023
- 2. The IoT revolution • “How Much Data Do We Create Every Day?” – Bernard Marr, Forbes, 21th May 2018 • Smart devices produce 5 quintillion (5 × 1018 ) bytes of data daily. • In 5 years, we can expect the number of these gadgets to be more than 50 billion! • 90 ZB (90 × 1021 bytes) of this data will be from IoT devices in 2025 • Response time? 2 SMART AGRICULTURE E-HEALTH FITNESS TRACKING TRAFFIC SAFETY Vincenzo De Maio - Distributed Systems in the Post-Moore Era
- 3. Traffic Safety • InTraSafEd5G Project • City of Vienna 5G Challenge • http://intrasafed.ec.tuwien.ac.at/ • Ensure traffic safety with the combination of IoT and Edge AI • Focus on near real-time performance • Need to consider users’ reaction time… Vincenzo De Maio - Distributed Systems in the Post-Moore Era 3
- 4. Cloud/Edge Offloading 4 Computationally Intensive Tasks App modeled as DAG Josip Zilic, Vincenzo de Maio, Atakan Aral, Ivona Brandic Edge offloading for microservice architectures. EdgeSys@EuroSys 2022: 1-6 RUCON LiveLab Testbed EDGE CLOUD Vincenzo De Maio - Distributed Systems in the Post-Moore Era
- 5. Edge infrastructure for traffic safety Vincenzo De Maio - Distributed Systems in the Post-Moore Era 5 Example setup and deployment of edge nodes in the context of InTraSafEd5G project. Ivan Lujic, Vincenzo De Maio, Klaus Pollhammer, Ivan Bodrozic, Josip Lasic, Ivona Brandic: Increasing Traffic Safety with Real-Time Edge Analytics and 5G. EdgeSys@EuroSys 2021: 19-24
- 6. Main Challenges 6 Computationaly Intensive Tasks App modeled as DAG RUCON LiveLab Testbed PLACEMENT PROVISIONING RELIABILITY ENERGY TRUST Vincenzo De Maio - Distributed Systems in the Post-Moore Era
- 7. Post-Moore’s Law Computing • To improve performance of current architectures, we need to reduce component size… • Component size: hitting the atom limit! • Time to consider alternative (post-Moore’s Law) forms of computing • Quantum mechanics: interactions at the subatomic level • Quantum Computing: development of computer based on the principles of quantum theory • Qubits, superposition, entanglement… Vincenzo De Maio - Distributed Systems in the Post-Moore Era 7
- 8. Known Quantum Speedup • Grover’s algorithm: 𝑂( 𝑛) vs 𝑂(𝑛) • Shor’s algorithm: Polynomial vs Exponential • Quantum ML • Bayesian Inference: quadratic • SVM: exponential • Reinforcement Learning: quadratic • “Machine Learning: Quantum vs Classical”, Tariq M. Khan et al., IEEE Access, November 2020 Vincenzo De Maio - Distributed Systems in the Post-Moore Era 8
- 9. Quantum Fundamentals Qubits • |𝛙⟩ = 𝛼0 0 + 𝛼1 1 , 𝛼0, 𝛼1 ∈ ℂ 𝛼0 2 + |𝛼1|2 = 1 • 𝟎 = 𝟏 𝟎 BLOCH SPHERE • |𝛙⟩ = 𝛼0 0 + 𝑒𝑖φ 𝛼1 1 , 𝛼0, 𝛼1, φ ∈ ℝ • θ, φ: spherical coordinates with radius = 1 • |𝛙⟩ = cos 𝜃 2 |0⟩ + 𝑒𝑖𝜑𝑠𝑖𝑛 𝜃 2 |1⟩ Probability of |𝛙⟩ = 0 Probability of |𝛙⟩ = 1
- 10. Quantum Computation • Quantum register: combination of n qubits • Classical register: 1 out of 2𝑛 values at a time • Quantum register: 2𝑛 values AT THE SAME TIME. (Quantum Parallelism) • Measurement returns a state 𝑖 with probability 𝛼𝑖 2 • Repeated execution • Most probable result → final result of the algorithm • Quantum algorithms goals: • Achieve a distribution such that • One correct result appears with high probability • More than one correct result appear with high and similar probability
- 11. Example of single qubit operation • Manipulation of Qubit is done by using specific operators (gates) • 𝑋 = 0 1 1 0 , Y = 0 −𝑖 𝑖 0 , 𝑍 = 1 0 0 −1 (Pauli gates) • 0 ∗ 𝑋 = 0 0 1 1 0 = 1 0 0 1 1 0 • 0∗1+1∗0 1∗1+0∗0 = 0 1 𝑞0 X +
- 12. State of the art of Quantum Systems • Noisy Intermediate Scale Quantum (NISQ) architectures Vincenzo De Maio - Distributed Systems in the Post-Moore Era 12 IBM Q Quantum System at Semicon West Quantum state preparation Measurement Classic hardware • Translation from classic input in quantum state • Quantum compilation (from source code to circuit) • Error correction • Limited number of qubits available • Higher execution time with respect to classic equivalent
- 13. Measurement • Schrödinger’s cat • Measuring the value of a qubit collapses the value in 0 or 1 respectively with probability 𝜶𝟎 𝟐 and |𝜶𝟏|𝟐 • Wavefunction collapse • Different measurements -> different results!
- 14. Notes • No-Cloning Theorem: • It is IMPOSSIBLE to clone a qubit. • Quantum Entanglement • Bell’s state: 1 2 (|00⟩ + |11⟩) • (EPR paradox) • Computation not involving entangled qubits can be performed with same efficiency on classical computing • To achieve exponential quantum speedup, you MUST exploit entanglement (Jozsa/Linden 2003) • Applications to quantum teleportation / communication
- 15. Quantum Error Correction • Challenges • Redundancy doesn’t work (no cloning) • Bit/phase flips • Wavefunction collapse • Main research lines • Quantum Redundancy (expansion of Hilbert space) • Stabilizer codes • Surface codes “Quantum Error Correction: An Introductory Guide”, Joschka Roffe
- 16. Hybrid Quantum Systems Vincenzo De Maio - Distributed Systems in the Post-Moore Era 16 Quantum tasks Classic tasks Workflow Management System Scientific Workflow User Quantum machine Classic HPC Mapper
- 17. Quantum computing for Distributed Scientific Applications • Data intensive • Natural 3D modelling of scientific problems • N-body • Particle physics • Many computation can benefit from quantum speedup • Approximate optimization • Eigenvalue calculation Vincenzo De Maio - Distributed Systems in the Post-Moore Era 17 IDEA: Accelerate specific tasks by means of quantum hardware
- 18. A Molecular Dynamics Use Case • Analyzing trajectories of backbone 𝐶𝛼 atoms of amino-acids segments • Identifying collective variables capturing molecular motions in a region of interest Vincenzo De Maio - Distributed Systems in the Post-Moore Era 18 Atom segments 𝐷 = 0 ⋯ 𝐷𝐼𝐽 ⋮ ⋱ ⋮ 𝐷𝐼𝐽 𝑇 ⋯ 0 Distance matrix Read trajectory file User input 𝐷𝑣 = 𝜆𝑣 Find maximum eigenvalue Which of these can exploit quantum advantage?
- 19. Application decomposition Vincenzo De Maio - Distributed Systems in the Post-Moore Era 19 Atom segments 0 ⋯ 𝐷𝐼𝐽 ⋮ ⋱ ⋮ 𝐷𝐼𝐽 𝑇 ⋯ 0 Distance matrix Read trajectory file User input 𝐷𝑣 = 𝜆𝑣 Find largest eigenvalue End Device Classic HPC Quantum machine
- 20. Distance Matrix Initialization • CSWAP TEST • Input: 𝜑 , |𝜓⟩, quantum states • Outputs an estimate of | 𝜓 𝜑 |2 Vincenzo De Maio - Distributed Systems in the Post-Moore Era 20
- 21. Example calculation of interatomic distance Vincenzo De Maio - Distributed Systems in the Post-Moore Era 21 Select amino- acids segments 𝑑00 ⋯ 𝑑02 ⋮ ⋱ ⋮ 𝑑20 ⋯ 𝑑22 𝑎2 𝑎1 𝑎0 𝑏2 𝑏1 𝑏0 Amplitude encoding CSWAP TEST 𝜑 |𝜓⟩
- 22. Hybrid Testbed Vincenzo De Maio - Distributed Systems in the Post-Moore Era 22 Workflow Management System User Molecular Dynamics Workflow Classic HPC ibm_lagos ibmq_jakarta ibmq_lima ibmq_manila
- 23. Results for interatomic distance • 100 pairs of random generated matrices • Segment sizes: 1,2,4,8,16 • MSE between classic and quantum result Vincenzo De Maio - Distributed Systems in the Post-Moore Era 23 Node ID Average MSE Variance ibmq_manila 0.2317 0.000199 ibmq_santiago 0.2832 0.000264 ibm_lagos 0.2249 0.000190 ibm_jakarta 0.2037 0.000149
- 24. Calculation of eigenvalues • Variational Quantum Eigensolver (VQE) • In quantum mechanics, a system of particles can be described as a Hamiltonian representing the energy of the system. • Finding minimum eigenvalue ≡ Finding Hamiltonian ground state Vincenzo De Maio - Distributed Systems in the Post-Moore Era 24 𝐻 Ψ(Θ) Calculate expectation value 𝜆𝜃 = ⟨𝜓 Θ 𝐻 𝜓 Θ ⟩ 𝜆𝑚𝑖𝑛 ≤ 𝜆𝜃 𝐶(Θ) 𝑚𝑖𝑛𝐶(Θ) Molecular system Parametrized quantum circuit
- 25. Mapping of VQE Vincenzo De Maio - Distributed Systems in the Post-Moore Era 25 Classic Machine Quantum Machine 𝐻 𝜆𝜃 = ⟨𝜓 Θ 𝐻 𝜓 Θ ⟩ Θ 𝐶(Θ) Optimizer
- 26. Hyperparameter setting in VQE Vincenzo De Maio - Distributed Systems in the Post-Moore Era 26 Optimizer? Hardware? PQC? Cost function? Termination condition? Hamiltonian?
- 27. Parametrized Quantum Circuits • Standard “well-known” circuits • Entanglement • Repetitions Vincenzo De Maio - Distributed Systems in the Post-Moore Era 27 SU2 Pauli Two Design Real Amplitudes Excitation Preserving
- 28. Optimizers • Optimizers affect convergence rate and error • We select three optimizers for our evaluation • COBYLA • SPSA • GRADIENT DESCENT Vincenzo De Maio - Distributed Systems in the Post-Moore Era 28
- 29. PQC vs Quantum Hardware • Width: amount of qubits required to represent input matrix (𝑛 ∙ 𝑛 = log 𝑛) • Error due to decoherence and quantum noise Vincenzo De Maio - Distributed Systems in the Post-Moore Era 29
- 30. PQC vs Entanglement • Entanglement: • LINEAR: 𝑞0 → 𝑞1 → … → 𝑞𝑛 • FULL: 𝑞0 → 𝑞1, 𝑞2, … , 𝑞𝑛 , 𝑞1 → 𝑞0, 𝑞2, … , 𝑞𝑛 , … , 𝑞𝑛 → 𝑞0, 𝑞1, … , 𝑞𝑛−1 • SCA: 𝑞0 → 𝑞2, 𝑞4 … , 𝑞𝑛 • CIRCULAR: 𝑞0 → 𝑞1 → … → 𝑞𝑛 → 𝑞0 Vincenzo De Maio - Distributed Systems in the Post-Moore Era 30
- 31. PQC vs Repetitions • Error due to decoherence and quantum noise increases with respect to repetitions • Error correction? Vincenzo De Maio - Distributed Systems in the Post-Moore Era 31
- 32. Results • VQE calculation using different hyperparameters • Benchmarking data collected on different machines • Hyperparameters’ optimization is used to identify best hyperparameters set for a target metric 𝑚, Π𝑚 ∗ Vincenzo De Maio - Distributed Systems in the Post-Moore Era 32
- 33. Remarks • We provided a first step in the design of scientific applications for hybrid classic/quantum systems • Identified quantum-suitable parts • Provided an example implementation • Future work • Consider different use cases • Investigating impact of different quantum hardware • (semiconductors, ion-traps, d-wave…) • Error correction methods Vincenzo De Maio - Distributed Systems in the Post-Moore Era 33
- 34. Current Work 34 if not backend.configuration().simulator: trans_dict = {'layout_method': 'sabre', 'routing_method': 'sabre'} trans_circ = transpile(ansatz, backend, optimization_level=3, **trans_dict) vqe_inputs = { 'ansatz': trans_circ, 'shots': 8192, 'measurement_error_mitigation': True } options = { 'backend_name': backend.name(), } job = provider.runtime.run(program_id='vqe', inputs=vqe_inputs, options=options) MD Simulation Classic Code Quantum Circuit TRANSPILE Vincenzo De Maio - Distributed Systems in the Post-Moore Era DATA • Vincenzo De Maio, Atakan Aral, Ivona Brandic: A Roadmap To Post-Moore Era for Distributed Systems. ACM ApPLIED@PODC 2022: 30-34 • Sandeep Suresh Cranganore, Vincenzo De Maio, Tu Mai Anh Do, Ivona Brandic, Ewa Deelman: Molecular Dynamics Workflow Decomposition for Hybrid Classic/Quantum Systems. IEEE eScience 2022
- 35. Future development • Integration of different applications • Streaming data encoding • Quantum software engineering • … Vincenzo De Maio - Distributed Systems in the Post-Moore Era 35
- 36. Questions? Dr. Vincenzo De Maio vincenzo@ec.tuwien.ac.at