[OEP 9] Use 16k pages for next versions of OrientDB by default

[andrii0lomakin]

Reference:

https://github.com/orientechnologies/orientdb-labs/blob/master/OEP_9.md

Summary:

Use 16k pages for next versions of OrientDB by default

Goals:

Improve speed of read/writes operation by decreasing of default size of database page

Non-Goals:

Success metrics:

Improve numbers of performance benchmarks

Motivation:

If we decrease default size of cache page we decrease amplification of read and write operations.

Description:

According to our investigation during the database load:

1. Most of the non-leaf index pages are placed into the disk cache, and leaf pages are accessed using random IO pattern.
2. The difference between random IO and sequential IO much smaller on modern SSDs and in the case of big pages like 64KB time is needed to write page can be comparable with random access overhead.
3. All writes which we perform on the database are performed using sequential operations if possible.
4. The smaller size of pages the better efficiency of disk cache.

All of this suggest do decrease page size from 64KB which is huge to 16 KB, which should increase a speed of random IO operations and should not affect or should improve the speed of write operations.

Let's check all those speculations by real YCSB benchmark. If we run following YCSB loads (mixed load (50% writes, 50% reads), mostly read load (5% writes, 95% reads), read only load (100% of reads)) in case page sizes are 16 KB and 64 KB we get following results:

Load of data

64k pages

[INSERT], AverageLatency(us), 2659.181693165
[INSERT], MinLatency(us), 36.0
[INSERT], 95thPercentileLatency(us), 1935.0
[INSERT], 99thPercentileLatency(us), 3647.0

Total time: 18 hrs 32 mins 59.659 secs = 66 721 secs

16k pages

[INSERT], AverageLatency(us), 1649.60618966
[INSERT], MinLatency(us), 35.0
[INSERT], 95thPercentileLatency(us), 1770.0
[INSERT], 99thPercentileLatency(us), 2481.0

Total time: 11 hrs 31 mins 58.591 secs = 41 461 secs

Performance gain - 16k pages 1.61 times faster on load

Mixed load 50% of writes, 50% of reads

64k pages

[READ], AverageLatency(us), 2289.592112263472
[READ], MinLatency(us), 17.0
[READ], 95thPercentileLatency(us), 1616.0
[READ], 99thPercentileLatency(us), 18895.0

Thoroughput = 3494 op/secs

[UPDATE], AverageLatency(us), 7351.444064681804
[UPDATE], MinLatency(us), 36.0
[UPDATE], 95thPercentileLatency(us), 3161.0
[UPDATE], 99thPercentileLatency(us), 72191.0

Thoroughput = 1 088 op/secs

16k pages

[READ], AverageLatency(us), 1199.684215890254
[READ], MinLatency(us), 17.0
[READ], 95thPercentileLatency(us), 787.0
[READ], 99thPercentileLatency(us), 1353.0

Thoroughput = 6668 op/secs

[UPDATE], AverageLatency(us), 5039.77029837364
[UPDATE], MinLatency(us), 32.0
[UPDATE], 95thPercentileLatency(us), 1308.0
[UPDATE], 99thPercentileLatency(us), 2781.0

Thoroughput = 1 587 op/secs

Performance gain - read throughput is improved at 1.9 times, read latency is improved at 14 times !, write throughput is improved at 1.45 times, write latency is improved at 26 times !

Read mostly load 5% writes, 95% reads

64k pages

[UPDATE], AverageLatency(us), 2338.448307545592
[UPDATE], MinLatency(us), 41.0
[UPDATE], 95thPercentileLatency(us), 2319.0
[UPDATE], 99thPercentileLatency(us), 7059.0

Thoroughput = 3 421 op/secs

[READ], AverageLatency(us), 743.0695396303125
[READ], MinLatency(us), 16.0
[READ], 95thPercentileLatency(us), 1407.0
[READ], 99thPercentileLatency(us), 3575.0

Thoroughput = 10 767 op/secs

16k pages

[UPDATE], AverageLatency(us), 2492.1606831810336
[UPDATE], MinLatency(us), 44.0
[UPDATE], 95thPercentileLatency(us), 1049.0
[UPDATE], 99thPercentileLatency(us), 2615.0

Thoroughput = 3 210 op/secs

[READ], AverageLatency(us), 559.9084873968899
[READ], MinLatency(us), 18.0
[READ], 95thPercentileLatency(us), 668.0
[READ], 99thPercentileLatency(us), 1096.0

Thoroughput = 14 288 op/secs

Performance gain: write throughput is almost the same, write latency is improved at 2.6 times. Read throughput 1.32 times better, read latency is improved at 3.26 times.

Read only workload

64k pages

[READ], AverageLatency(us), 483.73006588
[READ], MinLatency(us), 16.0
[READ], 95thPercentileLatency(us), 1252.0
[READ], 99thPercentileLatency(us), 1825.0

Thoroughput = 16 538 op/secs

16k pages

[READ], AverageLatency(us), 263.79829212
[READ], MinLatency(us), 16.0
[READ], 95thPercentileLatency(us), 583.0
[READ], 99thPercentileLatency(us), 823.0

Thoroughput = 30 327 op/secs

Performance gain - read throughput 1.8 times better, read latency is improved at 2.2 times

Alternatives:

Risks and assumptions:

Maximum key size will be decreased from 10240 to about 5120

Impact matrix

• [X] Storage engine
• [ ] SQL
• [ ] Protocols
• [ ] Indexes
• [ ] Console
• [ ] Java API
• [ ] Geospatial
• [ ] Lucene
• [ ] Security
• [ ] Hooks
• [ ] EE

[lvca]

It looks like this work was worth it! Good job @laa.

[smolinari]

Yes. Absolutely!!! Being web based content is basically the 5/95 scenario, getting about 3 times more performance is brilliant! Great work! :+1:

Scott

[tglman]

Hi,

a +1 also from me, I think we should target this for 3.0.

Bye

[andrii0lomakin]

Hi @tglman I also thought about 3.0 but let discuss it closer to 3.0 release.