Amazon EC2 Ireland

Benchmark Instance

Web-Fronted

Benchmark Instance

Web-Fronted

...OS: Windows ServerSize: 1 core, 1 GB RAM

50 parallel instances

Benchmark Instance

Web-Fronted

Launched simultaneously

IIIInnttterrnnnneeeeettttInternetSquid Web

Cache

OS: LinuxSize: 2 cores, 7.5 GB RAM

Ore

stes Versant

Object Database

OS: Windows ServerSize: 2 cores, 7.5 GB RAM

Laun

cher

165ms Amazon EC2 California

Orestes: a REST protocol for horizontally scalable cloud database access

Observations

Reduction of server load Low latency

IIIIIIInnnnttttteeeeeerrrrrnnnnnnnneeeeeeeeetttttttttttInternet OrestesCache

Arrivingrequests

Requests reaching the server

IIIIIIInnnttttttteeeeeerrrrrnnnnnnnnnneeeeeeeettttttttttttInternetOrestes Cache

Low latency High latency

Orestes

Arbitrary combinations of persistence API and DB Reads scale horizontally

JavaScript JPA Orestes

Open protocol

Orestes

HTTP and JSON

VOD

Java JDO Orestes Orestes db4o

Different database systemsDifferent programming languages / APIs

IIIIIIInnnntttttteeeeeerrrrrnnnnnnnneeeeeeeeettttttttttttInternet OrestesLoad-Balancer

Arrivingrequests Scale-out

B

CacheCacheCacheCache writes

changes

Reads don't reach the server

PPPPPPPPPaaaaaaaaaaaaaaaaaaSSSSSSSSSSSCCCCClllooooooouuuuuuuuddddddddddddddPaaSCloud

CCCCCCllllllooooouuuuuuuuddddddddddddCloudOrestes as a

ServiceO

Orestes ApplicationOrestes

Application deployed in cloud

On premise application

Firewalls, proxys, load-balancers

AAAAAAAAAAA

Cloud and service capable DB interface

Browser persistence and administration

IIIIIIIInnnntttttteeeeeerrrrnnnnnnnneeeeeeeeettttttttttttInternet Orestes

Remote database

Browser

Accounts, data analysis, monitoring...

Direct database access from client side

s

Solution

Open problems: Database as a service, cloud computing

network latency Web-scale applications, commodity hardware

scalability issues Goal: Scalable, cloud capable database systems Our approach: ORESTES (Objects RESTfully encapsulated in standard formats)

Property Mechanism

Read scalability Web caching of database objects

Low latency Cache deployment in client network

Loose coupling of persistence API and DB

Generic HTTP access protocol and resource structure

Standard formats

Extensible HTTP content negotiation + default JSON formats

Felix Gessert, Florian Bücklers, Norbert Ritter University of Hamburg

Illustration of the Key Properties

System Overview

Caching mechanism: HTTP caching with statically assigned object lifetimes

Server

Persistence API

OrestesInterface RESTful HTTP

Orestes Protocol

Database Interface

OrestesClient

D bPersistence

Client

OrestesServer

OrestesInterface

Application Database

Scales horizontally

Orestes layer

Visit us at the URL http://orestes.info or contact us via email: research@orestes.info

Orestesserver(America)

caches(Hamburg)

benchmarkclient(Hamburg) 1 2 3

123

Orestes without cache

VOD native

Orestes with cacheOrestesO

IIIIIIIIIIInnnnnnnnnnnttttttttteeeeeeeeeeeeeeerrrrrrrrrrrrnnnnnnnnnnnnnnneeeeeeeeeeeeeeeeeeeetttttttttttInternet

LL2-JDOCaaache

L2-JDOCache

Different forward proxy caches

Round-trip time 180ms ± 5ms

Compared benchmark setups

Squid 2, Squid 3, Microsoft TMG

Benchmark procedure: starting with cold cache, performing 50/500 operations in 3 consecutive runs, using a read/write ratio of 50/90/99% and a working set size of 300 objects

Benchmarking the Protocol

Further Information

Single Client Scenario

50 100 150 200 250

300 / 1300 / 2300 / 3

3000 / 13000 / 23000 / 3

30000 / 130000 / 230000 / 3

Time [s]

Setu

p [#

obje

cts/

run]

VOD nativeOrestes

0 100 200 300 400 500 600

O: Squid 2O: Squid 3

O: TMGO: no cacheVOD: native

Latency [ms]

Setu

p

Cloud Computing Scenario Future Work

Conclusions

00 12550

Other (e.g. query)ReadsWrite

Cache is warming up

1. Run

2. Run

3. Run

100

O: Squid 2

O: Squid 3

O: Squid3 pa.

O: TMG

O: no cache

VOD: native

Ops: 500, Reads: 99%, 3. Run

Setu

p

450 reads/50 writes over 3 runs

Read-heavy scenario, warm cache

Execution Time [s] Execution Time [s]

VOD native Orestes

Preventing stale reads and transaction aborts using bloom filters:

ClientOrestesServer

Database

PUT/POST ObjectWritexWritex-1..Writey

Write-Lognow

TTL days ago

1 0 0 1 1 0 1 1h1

hk

...

foreach(OID)

OID

bit vector(Bloom filter)

hash functions

GET Writes since TTL

set bit to 1

Network Application

Bloo

m-F

ilter

IIInntttteeeerrrrnneeeeetttaannddd CCCaaaccchhheeessss

Internetand Caches

Writ

e-Lo

ggin

g

isStale(OID)?

1 0 0 1 1 0 1 1

1 0 0 1 1 0 1 1

contains?

Make O(1) check before each GET

if necessary: revalidate

Average execution time of 50 concurrent clients

Benchmark procedure: starting with cold cache, 500 operations, 3 consecutive runs, read/write ratio of 90%, 300/3000/30000 total database objects, 50 concurrent clients

Further caching support: server-side invalidation for content delivery networks and server-controlled reverse proxy caches.

We achieved: • Read scalability through web caching• Considerable reduction of network latency • Increase of read performance through caching of

working sets Orestes: a novel approach to achieve read scalability for transactional object oriented persistenceOpen problem: write scalability

Effects of cache hits: reducing the latency of object access by orders of magnitude

local in-memory cache hits

cache hits cache misses

Scenario social networking OO model inheritance, aggregation, etc. Access pattern transactional navigation Database Versant Object Database (VOD) Persistence API Java Data Objects (JDO) Protocols VOD TCP, Orestes Caches JDO L2 (VOD), web caches (Orestes) Concurrency single client, 50 parallel clients

DatabaseGeneric Orestes HTTP Protocol Internet

Java Data Objects

Java Persistence

API Port

pm.makePersistent(obj)

JavaScript

Java

pm.getObjectById(id)

pm.newQuery(qry)

em.persist(obj)

em.find(cls,id)

em.createQuery(qry,cls)

HTTP

Persistence API Generic Orestes HTTP Protocol

HTTP

Versant Object

Database

db4o

Bind

ing

Bind

ing

Optional web cache in client's network (forward proxy cache)

Optional web cache in servers's network (reverse proxy cache)

Web Cache

Web Cache Cluster

GET /db/sim.classes/Person/22517998If-None-Match: "2"Cache-Control: max-age=0Accept: application/json

Works in web browser (AJAX) and Node.js

Caches at ISPs

HTTP requestHTTP response

{ "_objectInfo": { "class": "/db/sim.classes/Person", "oid": "/db/sim.classes/Person/22517998", "version": "2" }, "/db/sim.classes/Site": { "owner": "/db/sim.classes/Person/22517998", "title": "Profile of Bethold Birnenbaum", "posts": [], "design": "Colorful" }, "/db/sim.classes/Person": { "status": "Chilling!", "family": [], "friends": [], "profile": "/db/sim.classes/Profile/22517998", "games": [] }}

200 OKCache-Control: public,max-age=1209600Content-Type: application/jsonETag: "2"

Content Delivery Networks (CDN)

Synchronization: Optimistic concurrency control to deal with stale objects

Implementation: Generic network layer, 2 persistence APIs, 2 DB bindings