Re: Large PostgreSQL servers

On Wed, Mar 21, 2012 at 3:31 PM, Kjetil Nygård <polpot78(at)gmail(dot)com> wrote:
> We wonder if someone could give some hardware / configuration specs for
> large PostgreSQL installations.
> We're interested in:
>        - Number of CPUs
>        - Memory on the server
>        - shared_buffers
>        - Size of the database on disk

Just yesterday I purchased this system:

Details:
CPU: 2 x Opteron 6212 (2.6GHz, 8-Core, G34, 16MB L3 Cache) 115W TDP, 32nm
RAM: 32GB (8 x 4GB) Operating at 1333MHz Max (DDR3-1333 ECC
Registered DIMMs)
NIC: Intel 82576 Dual-Port Gigabit Ethernet Controller - Integrated
Management: Integrated IPMI 2.0 & KVM with Dedicated LAN
PCIe 2.0 x8 - 1: LSI 9265-8i 6Gb/s SAS/SATA RAID (8-Port Int) with
1GB DDR3 Cache (BBU Consumes 2nd PCI Slot)
PCIe 2.0 x8 - 2: No Item Selected
NOTE: SAS Drives or More Than 6 SATA Drives Require Controller
(See PCIe 2.0 Slot)
Hot-Swap Drive - 1: 40GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 2: 40GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 3: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 4: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 5: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 6: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 7: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Hot-Swap Drive - 8: 160GB Intel 320 Series MLC (3Gb/s) 2.5" SATA SSD
Optical Drive: Low-Profile DVD-ROM Drive
Power Supply: Redundant 700W Power Supply with PMBus - 80 PLUS
Gold Certified
**** Additional Components ****
Fastpath: LSI FastPath Software License (Requires 926x 928x controller)

drives 1+2 RAID 1 for boot + OS
drives 3+4 RAID 1 for db logs
drives 5-8 RAID 10 for data

The expected maximum size of the stored data is going to be about
100GB, so the goal was to have the data area about 2x that limit since
they are SSDs and work optimally with lots of room to spare.

My currently largest server has 22GB RAM and I specify 5GB as shared
buffers. Most important in tuning is to get your random_page_cost
right, and have enough checkpoint segments for your write load.