Document in text mode:

Zoolander: Efficiently Meeting Very Strict, Low-Latency SLOsChristopher Stewart and Aniket ChakrabartiThe Ohio State UniversityRean GriffithVMWAREFunded in part by NSF EAGER 123077 Slide 2●For Internet services, slow response times cost100-400ms delay reduces searches per session [Google '09]100ms delay drops revenue by 1% [Crocker et al. '12]●Revenue >> Hardware Costs●To profit: Revenue > Hardware + Salaries + Benefits etc.Cost Effective Scaling 1. Traditional Scale Out 2. Big Data Challenges 3. The Time has Come 4. Contributions Slide 3●As arrival rate grows, processing tiers scale out●As data grows, data tiers scale out ●In big-data era, frequent data access per request–TripAdvisor: each request causes 20-40 memcached accesses [Gelfond, 2011]–Map-reduce services and graph processing issue 103--105Front EndApacheBusiness logicJboss, WebsphereData LookupMemcached, Redis,CassandraData ArchivesHadoop, TapeScales with: Request arrivalsData sizeCost Effective Scaling 1. Traditional Scale ...