This is a scratch pad for some PostgreSQL 8.0 benchmarks. The contributed utility pgbench is used for the testing.

For most of the testing, important parts of the PostgreSQL configuration used are

shared_buffers    = 23987
max_fsm_relations = 5950
max_fsm_pages     = 3207435

wal_buffers         = 544
checkpoint_segments = 40
checkpoint_timeout  = 900
checkpoint_warning  = 300
commit_delay        = 20000
commit_siblings     = 3
wal_sync_method     = fdatasync

enable_seqscan        = off
default_with_oids     = off
stats_start_collector = false

Exceptions are noted as the tests are performed.

The pgbench test database was created with the -s600 scale factor option. This results in a fresh database of about 8.6GiB, along with 1.3GiB of WAL. The test database was then backed up to a .tar.gz file so it could easily be restored between test runs.

Each test was executed 5 times in sequence, and the median result is reported. All tests were executed with the -c100 option for 100 connections. The transaction count per connection was adjusted as necessary so that each single test would span several minutes. Typical settings were -t500 to -t1000.

The pgbench client was actually run over a 100Mbit, full-duplex network connection from a client machine for all of the testing. Running pgbench remotely has not measurably degraded the performance. The client machine is a dual 3.06GHz Xeon running Linux 2.4.27. SSL encryption was disabled.

The base hardware:

• HP DL380 G4
• Dual 3.20GHz Xeon, 1MB L2 Cache, 800MHz FSB, HyperThreading disabled
• 1GB DDR2-400 (PC2-3200) registered ECC memory
• Broadcom PCI-X onboard NIC
• SmartArray 6i onboard RAID controller
• Battery-backed write cache enabled

The base software:

• LinuxKernel 2.4.27 from Debian’s kernel-image-2.4.27-2-686-smp Package
• Using Ext3 with 128MB journal and ordered data mode

On with the testing!

Update: On Tue Jun 20 2006, all results were replaced with updated results. The previous test results were invalid and incomparable, due to inconsistencies and errors in the testing process.

Results: 4-disk configurations

• Data array: RAID5, 4x 72GB 10k RPM
  WAL array: On data array
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 500
  number of transactions actually processed: 50000/50000
  tps = 124.728272 (including connections establishing)
  tps = 124.885813 (excluding connections establishing)

• Data array: RAID5, 4x 72GB 10k RPM
  WAL array: On data array
  Other notes: commit_delay disabled
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 500
  number of transactions actually processed: 50000/50000
  tps = 129.347747 (including connections establishing)
  tps = 129.517978 (excluding connections establishing)

• Data array: RAID5, 4x 72GB 10k RPM
  WAL array: On data array
  Other notes: battery-backed write cache disabled
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 500
  number of transactions actually processed: 50000/50000
  tps = 114.885220 (including connections establishing)
  tps = 115.020971 (excluding connections establishing)

• Data array: RAID5, 4x 72GB 10k RPM
  WAL array: On data array
  Other notes: Battery-backed write cache and commit_delay disabled
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 500
  number of transactions actually processed: 50000/50000
  tps = 80.177806 (including connections establishing)
  tps = 80.244181 (excluding connections establishing)

• Data array: RAID1+0, 4x 72GB 10k RPM
  WAL array: On data array
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 216.341138 (including connections establishing)
  tps = 216.606298 (excluding connections establishing)

• Data array: RAID1, 2x 72GB 10k RPM
  WAL array: RAID1, 2x 72GB 10k RPM
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 131.213838 (including connections establishing)
  tps = 131.309052 (excluding connections establishing)

• Data array: RAID1+0, 4x 72GB 15k RPM
  WAL array: On data array
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 284.662951 (including connections establishing)
  tps = 285.127666 (excluding connections establishing)

• Data array: RAID5, 4x 72GB 15k RPM
  WAL array: On data array
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 189.203382 (including connections establishing)
  tps = 189.379783 (excluding connections establishing)

• Data array: RAID1, 2x 72GB 15k RPM
  WAL array: RAID1, 2x 72GB 15k RPM
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 171.537230 (including connections establishing)
  tps = 171.680858 (excluding connections establishing)

Results: 6-disk configurations

• Data array: RAID1+0, 4x 72GB 15k RPM
  WAL array: RAID1, 2x 72GB 10k RPM
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 340.756686 (including connections establishing)
  tps = 341.404543 (excluding connections establishing)

• Data array: RAID5, 4x 72GB 15k RPM
  WAL array: RAID1, 2x 72GB 10k RPM
  

  scaling factor: 600
  number of clients: 100
  number of transactions per client: 1000
  number of transactions actually processed: 100000/100000
  tps = 212.377629 (including connections establishing)
  tps = 212.615105 (excluding connections establishing)

• Data array:
  WAL array:
  Other notes:
  

Insights and observations

• Using RAID1+0 for the heap files provides a dramatic performance gain. This is because RAID1+0 performs random write very well, compared to RAID5. Using RAID1+0 is an expensive option because of the additional disks (perhaps expensive SCSI disks), space, power and cooling capacity required. Whether the gain is worth the cost is a deployment specific question.

• Current PostgreSQL myths claim that moving the WAL to its seperate spindles, often on RAID1 or RAID1+0, increases performance. Most of the performance gain arises from increasing the number spindles the total IO load is distributed over, rather than the specific disk configuration. In particular it should be noted that:

  1. RAID5, with the help of a battery backed write cache, does sequential write very well.
  2. The WAL is written sequentially.
• The commit_delay parameter is a help for non-battery-backed systems, and a loss for battery-backed systems. The reason is clear: with a battery, the fsync() is almost free, thus the delay means the total throughput goes down. However without a battery the fsync() is very expensive, and if the delay allows it to be eliminated in the majority of cases, the throughput can go up.

pgbench test limitations

• The test is of the saturated throughput, rather than latency. In typical database usage, however, it is the latency that is the dominant performance metric. Many database situations have a cirtain bound on user acceptable latency. It may be interesting to perform throughput-vs-connections-vs-latency comparisons. Under such a test, the true gain of a battery-backed cache, and the true cost of commit_delay would be evident.
• The probability distrubution used to access the data is a completely flat probability curve. That is, all data items have the same probability of being accessed. This is completely unrepresentivive of typical situations where some data items are accessed very frequently and some hardly at all. Unfortunately the actual specific probability distribtuion is likely to be very application specific.
• The TPC-B benchmark, which the pgbench program is somewhat based upon, was actually retired by the Transaction Processing Council (TPC) in 1995. The test database schema is very basic, with very simple queries. Better benchmark tools are available, such as the DBT suite developed by Open Source Development Labs (OSDL) which aim to be a fair-use implementation of the current TPC benchmarks.

Other observations

• The WAL consumes large amounts of Kernel page cache. When moving the WAL between devices, when the old files are unlinked, 1/2 of the page cache is freed. Since the WAL is never read and written only once, this is as waste!
• The battery-backed write cache makes write performance very erratic.
• The HP SmartArray hardware (or perhaps driver) tends to block reads while there are cached writes occuring. Large read latencies (whole seconds) result. I have not yet found a way to tune this.

Part of CategoryDiskNotes