Servers
Slide #1
Topics
1. Servers vs Desktops
2. Server Hardware
3. Operating System
4. Different Approaches to Servers
Slide #2
How are Servers different?
• 1000s of clients depend on server.
• Requires high reliability.
• Requires tighter security.
• Often expected to last longer.
• Investment amortized over many clients,
longer lifetime.
Slide #3
Vendor Product Lines
Home
– Cheapest purchase price.
– Components change regularly based on cost.
Business
– Focuses on Total Cost of Ownership (TCO).
– Slower hardware changes, longer lifetime.
Server
– Lowest cost per performance metric (nfs, web)
– Easy to service rack-mountable chassis.
– Higher quality (MIL-SPEC) components.
Slide #4
Server Hardware
• More internal space.
• More CPU/Memory.
– More / high-end CPUs.
– More / faster memory.
• High performance I/O.
– PCIe vs PCI
– SCSI/FC-AL vs. IDE
• Rack mounted.
• Redundancy
– RAID
– Hot-swap, hot-spares
Slide #5
Rack Mounting
Efficient space utilization.
Simple, rectangular shape measured in RUs.
Repair and upgrade while mounted in rack.
No side access required.
Requirements
Cooling through back, not sides.
Drives in front, cables in back.
Remote management (serial console, hardware
sensors, VM MUI)
Slide #6
Server Memory
Servers need more RAM than desktops.
– x86 supports up to 64GB with PAE.
– x86-64 supports 1 PB (1024 TB)
Servers need faster RAM than desktops.
– Higher memory speeds.
– Multiple DIMMs accessed in parallel.
– Larger CPU caches.
Slide #7
Server CPUs
Intel Xeon
• Up to 8 cores with 2 threads each @ 1.8 to 3.3 GHz
• Up to 18 MB L3 cache
AMD Opteron
• 4, 6, 8, or 12 cores @ 1.4 to 3.2 GHz
• Up to 12 MB L3 cache
IBM Power 7
• 4, 6, or 8 cores with 4 threads each @ 3.0 to 4.25 GHz
• 4 MB L3 cache per core (up to 32MB for 8-core)
Sun Niagara 3
• 16 cores with 8 threads each @ 1.67 GHz
• 6 MB L2 cache
Slide #8
Xeon vs Pentium/Core CPUs
Xeon based on Pentium/Core with
changes that vary by model:
– Allows more CPUs
– Has more cores
– Better hyperthreading
– Faster/larger CPU caches
– Faster/larger RAM support
Slide #9
System Buses
Servers need high I/O throughput.
– Fast peripherals: SCSI-3, Gigabit ethernet
– Often use multiple and/or faster buses.
PCI
– Desktop: 32-bit 33 MHz, 133 MB/s
– Server: 64-bit 66 MHz, 533 MB/s
PCI-X (backward compatible)
– v1.0: 64-bit 133 MHz, 1.06 GB/s
– v2.0: 64-bit 533 MHz, 4.3 GB/s
PCI Express (PCIe)
– Serial architecture, v3.0 up to 16 GB/s
Slide #10
Hardware Redundancy
Disks are most likely component to fail.
– Use RAID for disk redundancy.
– Cover in detail in Disks lecture.
Power supplies second most likely to fail.
– Use redundant power supplies.
– Many servers need 2 power supplies
normally.
– Need 3 power supplies for redundancy.
– Use separate power cord and UPS for each
power supply.
Slide #11
Full and n+1 Redundancy
n+1 Redundancy: One component can fail, but the system is
still functional.
– Ex: RAID 5, dual NICs with failover
Full Redundancy: Two complete sets of hardware configured
with failover mechanism.
– Manual: SA switches to 2nd system when notices failure.
– Automatic: The second system monitors the first and switches over
automatically on failure.
– Load-sharing: Both systems serve users, sharing load, but each has
capacity to handle entire load on its own. When one fails, other
automatically handles entire load.
Slide #12
Hot-swap Components
Hot-swap components
– Components can be replaced while running.
– Need n+1 redundancy for this to be useful.
– Don’t need to schedule a downtime.
Issues
– Which parts are hot-swappable?
– May require a few seconds to reconfigure.
– Be sure components are hot-swap, not hot-plug.
Slide #13
Hot Plug and Hot Spare
Hot Plug
– Electrically safe to replace component.
– Part may not be recognized until next reboot.
– Requires downtime, unlike hot swap.
Hot Spare
– Spare component already plugged into system.
– System automatically uses hot spare when disk/CPU
board etc. fails.
– Provides n+2 redundancy.
Slide #14
Separate Administrative Network
Reliability
– Allows access to machines even when network
is down.
Performance
– Backups require so much bandwidth that they’re
often done over their own network.
Security
– Network security monitoring data and logs sent
across network should be secured.
Slide #15
Maintenance Contracts
• All machines eventually break.
• Vendors offer variety of maint contracts.
• Non-critical: Next-day or 2-day contract.
• Clusters: If you have many similar hosts (CPU or web farm),
then on-site spares may be cheaper than maintenance contract.
• Controlled Model: Use small # of machine types for all
servers, so you can afford a spares kit.
• Critical Host: Same-day response or on-site spares.
• Highly Critical: On-site technician + dup machine.
Slide #16
Data Protection
• Avoid desktop backups by storing data on servers.
Easy on UNIX, harder on Windows.
• Use RAID for server hardware failures.
– Mirror root disk, higher RAID levels for data.
– Some servers use 16GB Flash drives for root disk.
– Doesn’t protect against software mistakes.
• Server backups
– Use specialized admin network to keep load off main
network.
– Use specialized tape jukeboxes to fully automate
backups of large data servers (DBs, fileservers).
Slide #17
Keep Servers in Data Center
Data center necessary for server reliability.
– Power (enough power, UPS)
– Climate control (temperature, humidity)
– Fire protection
– High-speed network
– Physical security
Slide #18
Server Operating Systems
Slide #19
Server OS Image
Need greater reliability, security than desktop.
– Remove unnecessary OS components.
– Configure for best security & performance.
Install and config specialized server software.
– Server software: web, db, nfs, dns, ldap, etc.
– May need monitoring software too.
– Configuration: disk space, networking
Server OS install should be automated too.
Slide #20
Remote Administration
Servers must be accessible remotely.
– Allows SA to fix problems quickly at 3am.
– Allows SA to work outside machine room.
Remote Administration
– Serial console and concentrator (UNIX)
– Networked KVM (Windows)
– Remote power control.
– Important to secure remote admin facilities.
Slide #21
Server Appliances
Dedicated hardware + software
– Fileserver (NetApp, Auspex)
– Print servers
– Routers
Advantages
– Performance
– Reliability
– Easy to setup
– Extra capabilities
Disadvantages
– Cost
Slide #22
Many Inexpensive Workstations
Why buy server hardware?
– Buy two cheap rack-
mount PCs + failover
software.
– Works if two PCs
cheaper than server.
– Google’s approach
with ~450,000 servers.
Slide #23
Blade Servers
• High-density
servers on a board.
– CPU
– Memory
– Disk
• Each blade lives in
a blade chassis.
Slide #24
Blade Chassis
• Blade chassis
provides power,
network, remote.
• Typically hot-
swappable, hot-spare.
• Racks can only
support 1 svr/RU.
• Blades are higher
density, but also
require more power
and cooling.
Slide #25
Key Points
Servers vs desktops
– Requirements and hardware differences.
Redundancy
– Full vs n+k redundancy.
– Hot plug vs hot spare.
Services
– Requirements: service, server, customer, operational.
– Machine independence and open architectures.
Performance
– Latency vs. throughput.
Slide #26
