Algorithms for Dispersed Processing Josef Spillner, Alexander Schill - PowerPoint PPT Presentation

Department of Computer Science | Institute of Systems Architecture | Chair of Computer Networks Algorithms for Dispersed Processing Josef Spillner, Alexander Schill mailto:josef.spillner@tu-dresden.de xmpp:josef.spillner@jabber.org 1 st International Workshop on Advances in Cloud Computing Legislation, Accountability, Security and Privacy (CLASP), December 8-11, 2014, London, UK

Background: Informatjon Dispersal Motivation »Don't trust the cloud.« Risks: ● temporary unavailability ● permanent unavailability ● of data [loss] ● of service [bankruptcy] ● arbitrary slowness ● unauthorised access [honest-but-curious] || 3rd-party spying ● malicious modification ● refusal to delete ● lock-in, transfer costs ● ... etc. pp. Insufficient but essential protection: Information dispersal over multiple clouds. => IDA 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 2 J. Spillner & A. Schill

Background: Informatjon Dispersal Definition Popular definition: [Image source: techtarget.com] But: Why just sit or traverse? Storage « √ Networking « √ Processing « ??? 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 3 J. Spillner & A. Schill

Background: Informatjon Dispersal Generalised Dispersed Processing Storage Network (Communication) Processing (Computation) data locality 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 4 J. Spillner & A. Schill

Dispersed Processing: Map-Reduce (a), (c): central processing (b), (d): map-reduce IDA of choice: bitsplitting; e.g. 50% redundancy: k=2, m=1, n=k+m Property: structure-preserving 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 5 J. Spillner & A. Schill

Dispersed Processing: Summatjon Algorithm Addition: requires n=2 Multiplication: requires n=4 (unless a, b ≠ 0) Method: - map: sum/mult - reduce: sum 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 6 J. Spillner & A. Schill

Dispersed Processing: Search Algorithm Pattern search without index Method: - split pattern, too - map: perform k partial searches - reduce: filter out false positives 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 7 J. Spillner & A. Schill

Dispersed Processing: Map-Carry-Reduce Limits of parallelisation - fixed-size integer shifts → carry bits - filter operations → position data Security effect: Information leakage 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 8 J. Spillner & A. Schill

Dispersed Processing: Encryptjon Processing dispersed + encrypted data blocks Cloud Speicher Speicher Source Quelle Quelle Verschlüsselung Encryption Splitting Cloud ⚷ ⚷ Ѧ Policies Property: structure-preserving! Algorithms: - homomorphic encryption → arithmetics - order-preserving encryption → sorting - convergent encryption → deduplication 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 9 J. Spillner & A. Schill

Dispersed Processing: Encryptjon Example generate-keypair bits=16 => {k priv , k pub } Travel distance: Arithmetic: a*b%bits 283060154 71 km 71 km 283060154 “km“ k pub + 19 km + 630596813 “km“ ----------- ----------- = 90 km 19 km = 540987952 “km“ 630596813 k pub k pub , k priv Travel distance: 540987952 „km“ = 90 km 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 10 J. Spillner & A. Schill

Dispersed Processing: Precision If a cloud fails... ... repair and continue processing? ... or, expect degraded results? Application-specific decision. Example: floating point with (i)nteger, (f)ractional, (r)edundant parts. 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 11 J. Spillner & A. Schill

Dispersed Processing: Redundancy Redundant data is not generally processable. Distribution matters. 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 12 J. Spillner & A. Schill

Dispersed Processing: Algorithms Overview Classification of algorithms (pending) 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 13 J. Spillner & A. Schill

Performance Evaluatjon 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 14 J. Spillner & A. Schill

Conclusion & Stealth Roadmap Dispersed Computing Storage, networking: Much existing research Processing: need for special algorithms Evaluation: slower processing vs. (often) less transmission Code: git://nubisave.org/git/dispersedalgorithms Stealth Computing Combination of dispersion x encryption and further quality measures Enabler for native cloud applications Stealth Apps Towards Towards for Secure Stealth Dispersed Cloud Dispersed Cloud Personal Data Databases Computing Computing Analytics ... [BSC'14] [CLASP'14] [NetSys'15] 1 st CLASP , 11.12.2014 Algorithms for Dispersed Processing # 15 J. Spillner & A. Schill