Hardware (The part you can kick)

Overview  Selection Process  Equipment Categories  Processors  Memory  Storage  Support

Define Requirements  Service Level Agreements  People  Process  Technology  Budget  You may get an AttaBoy for saving a few dollars.  You WILL get blamed for an inadequate system

What is the difference?  Workstation  Single user (desktop/notebook)  Server  Department  Enterprise

Processors  32 bit  64 bit  AMD  Intel  Single-Core  Multi-Core

32 Bit Processors  Proven, mature technology  Low Cost  Compatible  4GB Memory Address Limit  PAE  3GB  AWE  Called x86 (Intel was the first)

64 Bit Processors  Itanium  Incompatible with existing x86 processors  8 or more proc systems Enterprise Systems  AMD64/EMT64 (x64)  Pioneered by AMD. Intel dragged kicking and screaming into the market.  Binary compatible with x86 procs  Extensions for 64 bit commands  Very price effective (10%-15% premium)

Intel  Front Side Bus  Uniform memory access  1-4 processors is easy and fast  Potential memory bottlenecks  Decouples memory and CPU clock rates

AMD  AMD  NUMA NonUniform Memory Architecture  Memory directly controlled by processor  Matches SQLOS memory alignment  No bus bottlenecks  Processor action required for cross-boundary memory access  SQL 2005 native

AMD vs. Intel Drawing © AMD Technologies

Cores  Single-Core  Cheap, proven  Multi-core  Newer  Licensing advantage for SQL (sockets, not cores)  On-die  In-Package  HyperThreading  Simulates multiple cores. 10% to 15% performance boost for most SQL applications  Can slow down system with many non-parallelizable queries

Memory  Speed  Latency  Error Correction

Storage  Disk characteristics  Storage Subsystems  RAID Configuration

Disk Characteristics  RPM – Higher is better  On disk read cache – More is better  SCSI  Fast, reliable, proven, mature  SATA  Cheaper, slower, reliability? (new)  Good for second tier Storage (backup, historical partitions)  Fibre Channel

Storage Types  Locally Attached  SCSI  SATA  NAS  iSCSI  SAN

Local Storage  Hot swap disks  Controllers  Battery-backed cache  Disk arrays  Cache on controller  Cache in cabinet (clusterable)

NAS  Not a block-write device  SMB drive mapping  Cache corruption  Must be on Windows Catalog  CANNOT CLUSTER!!!

iSCSI  SCSI via IP  Lower Cost  Competes with NAS  Large variance in speed, quality, managability, reliability  Cluster nightmare

SAN  Pseudo-SAN – “Smart Array”  Large write cache (GB)  Fibre Channel (SCSI++)  Options  Split mirror backups  Distance synching  Virtual Snapshot  Management (Provisioning) tools

Disk Stripe Alignment  Low-level stripes (Smart Array controller or SAN)  OS-level stripes  MBR boot record offset (63 sectors)  Cache copying. IO doubling.  Up to 40% slowdown.  Diskpart.exe (Win2003 SP1) can pad offset to match stripe size  Diskpar.exe (Win 2000 Resource Kit) for older systems

RAID  Redundant Array of Inexpensive (later Independent) Disks  1988 ACM SIGMOD paper “A Case for Redundant Arrays of Inexpensive Disks (RAID)” by David A. Patterson, Garth A. Gibson, and Randy H. Katz.  Overcomes size and reliability limitations of physical disks

RAID Levels  JBOD Just A Bunch of Disks  RAID 0 Striping with no redundancy  RAID 1 Mirroring  RAID 2-3 not worth talking about  RAID 4 Stripe with Parity Disk  RAID 5 Stripe with Parity Rotation  RAID 1+0 (10) Stripe of mirrored pairs  RAID 0+1 (10) Mirror of stripes  RAID 50 Stripe of stripes w/ parity

RAID Performance  RAID 0  Read  Stripe advantage N  Write  1  RAID 1  Read  Stripe Advantage 2  Write  2x  RAID 5  Read  Stripe Advantage N-1  Write  N-2 reads + 2 Writes  RAID  Read  Stripe Advantage N  Write  2x

Baseline Performance  SQLIO stress utility  IOMeter stress utility  Performance Monitor capture

Support  Life cycle of equipment  Parts/service SLA  Management and Monitoring software  Working relationship

Questions