Computer Architecture Topic Explaination

 

Introduction

 

 

 

n          Warehouse-scale computer (WSC)

 

n       Provides Internet services

 

 

n    Search, social networking, online maps, video sharing, online shopping, email, cloud computing, etc.

 

n       Differences with HPC “clusters”:

 

n    Clusters have higher performance processors and network

 

n    Clusters emphasize thread-level parallelism, WSCs emphasize request-level parallelism

 

n       Differences with datacenters:

 

n    Datacenters consolidate different machines and software into one location

 

n    Datacenters emphasize virtual machines and hardware heterogeneity in order to serve varied customers

 

 

 

 

 

 

 

Introduction

 

 

 

n          Important design factors for WSC:

 

n       Cost-performance

 

n    Small savings add up

 

n       Energy efficiency

 

n    Affects power distribution and cooling

 

n    Work per joule

 

n       Dependability via redundancy

 

n       Network I/O

 

n       Interactive and batch processing workloads

 

n       Ample computational parallelism is not important

 

n    Most jobs are totally independent

 

n    “Request-level parallelism”

 

n       Operational costs count

 

 

 

n    Power consumption is a primary, not secondary, constraint when designing system

n       Scale and its opportunities and problems

 

n    Can afford to build customized systems since WSC require volume purchase

 

 

 

 

 

 

Program’s Models and Workloads

 

 

 

n          Batch processing framework: Map Reduce

 

 

 

n       Map: applies a programmer-supplied function to each logical input record

 

n    Runs on thousands of computers

 

n    Provides new set of key-value pairs as intermediate values

 

 

 

n       Reduce: collapses values using another programmer-supplied function

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Program’s Models and Workloads

 

 

 

n          Example:

 

n       map (String key, String value):

 

n    // key: document name

 

n    // value: document contents

 

n    for each word w in value

 

n    Emit Intermediate(w,”1”); // Produce list of all words

 

 

 

n       reduce (String key, Iterator values):

 

n    // key: a word

 

n    // value: a list of counts

 

n    int result = 0;

 

n    for each v in values:

 

n    result += ParseInt(v); // get integer from key-value pair

 

n    Emit(AsString(result));

 

 

 

 

 

 

Program’s Models and Workloads

 

 

 

n          MapReduce runtime environment schedules map and reduce task to WSC nodes

 

 

 

n          Availability:

 

n       Use replicas of data across different servers

 

n       Use relaxed consistency:

 

n    No need for all replicas to always agree

 

 

 

n          Workload demands

 

n       Often vary considerably

 

 

 

 

 

 Computer Architecture of WSC

 

 

 

n          WSC often use a hierarchy of networks for interconnection

 

n          Each 19” rack holds 48 1U servers connected to a rack switch

 

n          Rack switches are uplinked to switch higher in hierarchy

 

n       Uplink has 48 / n times lower bandwidth, where n = # of uplink ports

 

 

n    “Oversubscription”

 

n       Goal is to maximize locality of communication relative to the rack

 

 

 

 

 

 

 

 

 

Storage

 

 

 

n          Storage options:

 

n       Use disks inside the servers, or

 

n       Network attached storage through Infiniband

 

 

 

n       WSCs generally rely on local disks

 

n       Google File System (GFS) uses local disks and maintains at least three relicas

   

 

 

 Array Switch

 

 

 

n          Switch that connects an array of racks

 

n       Array switch should have 10 X the bisection bandwidth of rack switch

 

n       Cost of n-port switch grows as n2

 

n       Often utilize content addressible memory chips and FPGAs

 

WSC Memory Hierarchy

 

 

 

n          Servers can access DRAM and disks on other servers using a NUMA-style interface

 

 

Infrastructure and Costs of WSC

 

 

 

n          Location of WSC

 

n       Proximity to Internet backbones, electricity cost, property tax rates, low risk from earthquakes, floods, and hurricanes

 

n          Power distribution

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Infrastructure and Costs of WSC

 

 

 

n          Cooling

 

n       Air conditioning used to cool server room

 

n       64 F – 71 F

 

n    Keep temperature higher (closer to 71 F)

 

n       Cooling towers can also be used

 

n    Minimum temperature is “wet bulb temperature”

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Infrastructure and Costs of WSC

 

 

 

n          Cooling system also uses water (evaporation and spills)

 

n       E.g. 70,000 to 200,000 gallons per day for an 8 MW facility

 

 

n          Power cost breakdown:

 

n       Chillers: 30-50% of the power used by the IT equipment

 

n       Air conditioning: 10-20% of the IT power, mostly due to fans

 

 

n          How man servers can a WSC support?

 

n       Each server:

 

n    “Nameplate power rating” gives maximum power consumption

 

n    To get actual, measure power under actual workloads

 

n       Oversubscribe cumulative server power by 40%, but

 

Monitor power closely

 

 

 

 

 

 

 

 

 

 

Measuring Efficiency of a WSC

 

 

n          Power Utilization Effectiveness (PEU)

 

n       = Total facility power / IT equipment power

 

n       Median PUE on 2006 study was 1.69

 

 

 

n          Performance

 

n       Latency is important metric because it is seen by users

 

n       Bing study: users will use search less as response time increases

 

n       Service Level Objectives (SLOs)/Service Level Agreements (SLAs)

 

n    E.g. 99% of requests be below 100 ms

 

 

 

 

 

 

 

 

Cost of a WSC

 

 

 

n          Capital expenditures (CAPEX)

 

n       Cost to build a WSC

 

 

 

n          Operational expenditures (OPEX)

 

n       Cost to operate a WSC

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Cloud Computing

 

 

 

n          WSCs offer economies of scale that cannot be achieved with a datacenter:

 

n       5.7 times reduction in storage costs

 

n       7.1 times reduction in administrative costs

 

n       7.3 times reduction in networking costs

 

n       This has given rise to cloud services such as Amazon Web Services

 

n    “Utility Computing”

 

n    Based on using open source virtual machine and operating system software