Solid State File Cache

A New Infrastructure for Growing Your Business Michael CaseySolid Data Systemsmcasey@soliddata.comSession 9326

AgendaWhy File Cache?File Cache for Transaction ServersBenefits: Speed + Persistence in Real ApplicationsDiagnostic tools at CPU, I/O, & Application LevelsNew Applications = New InfrastructureTelecom Financial ServicesFile cache for Storage NetworksConclusions

Why File Cache? Waiting for I/O:Performance drops radically as users are addedSystem no longer able to keep upBatch jobs dont complete within the available processing time windowMonth-end close takes days instead of hoursExtensive tuning has not solved performance issuesSAR (Unix) or PerfMon (NT) indicate >50% I/O Wait

The Solution: Solid-State File CacheFile Cache

Bandwidth Bottleneck Transaction TraumaBandwidth(MB/second)NetworklimitedServerlimitedTransaction Rate (transactions/second)From Content to Commerce7

File Cache for Transaction Servers

File Cache vs Cached RAIDEntire file in RAM Latency as low as 0.014 ms Consistent performance No tuning required Best for small block random I/ONo moving parts - very high MTBFIndependently scalableSelected blocks in RAM - other blocks on diskLatency of 7.000 ms or greaterPerformance depends on cache algorithmsExtensive tuning sometimes requiredBest for large block streaming I/OMechanical complexity reduces MTBFLimited cache to RAID ratio

Speed and Persistence

Real-World Application BenefitsEmail server message capacity increased from 5/sec to over 40/secSatellite-based stock data recording backlog reduced from 4 hours to 10 millisecondsServer cluster fail-over reduced from 15 minutes to 7 minutesOverall file system speed increased by 25%Billing application reduced from 5 days to 2 days Batch job reduced from 8 hours to 2 hoursBatch job reduced from 72 hours to 8 hours System response time after data entry reduced from 15 seconds to 3 seconds

Diagnostic ToolsCPU utilization is the server I/O bound?Standard system tools SAR, Perfmon, etc. Third-party utilities e.g., I/O Dynamics* I/O profile: is it skewed, with hot files?Standard sys admin tools and displaysThird-party utilities e.g., I/O Dynamics* Application-level benchmark tests100% reads, 100% inserts, 100% updatesSynthetic workload generator

* Available for free download

CPU Utilization % Wait for I/OSource: Solid Data measurement of an e-mail system workload over 24-hour periodI/O Wait = NegligibleGeneral Purpose ServerTransaction Server% CPU LoadI/O Wait = 40% of CPU timeI/O WaitSystemUser12

I/O Activity vs. File Size

Real-Time Database Benchmark Source: Speed and Persistence for Real-Time Transactions July 2002 white paper www.soliddata.com/timesten

New Applications New InfrastructureRich functionality with real-time response6

Telecommunications: Enabling New Services for Revenue Growth

Wireless text messaging (SMS) Prepaid mobile billingAnti SpammingCall detail record (CDR) mediationFraud detection Instant Messaging

Ring-tone downloadLocation-based servicesMobile Internet (WAP)Multimedia Messaging Service (MMS) Micro-payments15

Case Study: Wireless Text Messaging

ProblemSuperior product offering (functionality) was not meeting throughput and response time requirements

File Cache-based Solution2 mirrored 10GB file caching systems

Business BenefitsHigher throughput at lower cost per subscriberAccelerated purge expands daily window Richer functionalityHigher reliability Improved scalabilityLower cost of management

16

SMS Application

10 GB

mirrored file cache

Short Messaging

Services Network

Wireless Network

ISP

Oracle or

Ingres

DB

HP K, L, N Class Servers

Financial Services: Real-Time TransactionsOnline securities tradingPortfolio risk managementMessaging middlewareReal-time databases

Credit card fraud detection T+1 Trade SettlementStraight Through Processing (STP)19

File Cache for Storage Networks

Tape / ATLLAN:SAN:Storage:Servers:Solid State File Cache

Shared File Cache in the SANE-Mail queues & server-to-server messaging queuesNon-volatile shared memory for server clustersFile system journalingSnap-shot device for backup / remote copyVirtualization lookup tables (address mapping)DatabasesRollback segments- Transaction logsTemp spaces- Hot indexesHot tablesSRM policy based management

ConclusionsThe Internet has pushed many applications to the limits of scalability, causing I/O bottlenecksSolid-state file cache is low latency storage complementary to cached RAIDSolid-state file cache is a cost efficient way to multiply performance and scalability of existing servers Diagnostic tools and benchmarks can help diagnose application issues and determine the fit for file cacheTransaction and messaging applications especially in telecom and financial services require a new infrastructureSAN connections and management software can extend the benefits of solid-state file cache to more servers and applications