Postgres 18rc1 vs sysbench

This post has results for Postgres 18rc1 vs sysbench on small and large servers. Results for Postgres 18beta3 are here for a small and large server.

tl;dr

• Postgres 18 looks great
• I continue to see small CPU regressions in Postgres 18 for range queries that don't do aggregation on low-concurrency workloads. I have yet to explain that. 
• The throughput for the scan microbenchmark has more variance with Postgres 18. I assume this is related to more or less work getting done by vacuum but I have yet to debug the root cause.

Builds, configuration and hardware

I compiled Postgres from source for versions 17.6, 18 beta3 and 18 rc1.

The servers are:

• small
• an ASUS ExpertCenter PN53 with AMD Ryzen 7735HS CPU, 32G of RAM, 8 cores with AMD SMT disabled, Ubuntu 24.04 and an NVMe device with ext4 and discard enabled.
• large32
• Dell Precision 7865 Tower Workstation with 1 socket, 128G RAM, AMD Ryzen Threadripper PRO 5975WX with 32 Cores and AMD SMT disabled, Ubuntu 24.04 and and NVMe device with ext4 and discard.
• large48
• an ax162s from Hetzner with an AMD EPYC 9454P 48-Core Processor with SMT disabled
• 2 Intel D7-P5520 NVMe storage devices with RAID 1 (3.8T each) using ext4
• 128G RAM
• Ubuntu 22.04 running the non-HWE kernel (5.5.0-118-generic)

All configurations use synchronous IO which is the the only option prior to Postgres 18 and for Postgres 18 the config file sets io_method=sync.

Configuration files:

• small server
• for versions 17.6 (x10a_c8r32) and 18 (x10b_c8r32)
• large servers
• for versions 17.6 (x10a_c32r128) and 18 (x10b_c32r128)

Benchmark

I used sysbench and my usage is explained here. To save time I only run 32 of the 42 microbenchmarks 

and most test only 1 type of SQL statement. Benchmarks are run with the database cached by Postgres.

For all servers the read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

The number of tables and rows per table was:

• small server - 1 table, 50M rows
• large servers - 8 tables, 10M rows per table

The number of clients (amount of concurrency) was:

• small server - 1
• large32 server - 24
• large48 servcer- 40

Results

The microbenchmarks are split into 4 groups -- 1 for point queries, 2 for range queries, 1 for writes. For the range query microbenchmarks, part 1 has queries that don't do aggregation while part 2 has queries that do aggregation. 

I provide charts below with relative QPS. The relative QPS is the following:

(QPS for some version) / (QPS for Postgres 17.6)

When the relative QPS is > 1 then some version is faster than PG 17.6.  When it is < 1 then there might be a regression. Values from iostat and vmstat divided by QPS are also provided here. These can help to explain why something is faster or slower because it shows how much HW is used per request.

The numbers highlighted in yellow below might be from a small regression for range queries that don't do aggregation. But note that this does reproduce for the full table scan microbenchmark (scan). I am not certain it is a regression as this might be from non-deterministic CPU overheads for read-heavy workloads that are run after vacuum. I hope to look at CPU flamegraphs soon.

Results: small server

I continue to see small (~3%) regressions in throughput for range queries without aggregation across Postgres 18 beta1, beta2, beta3 and rc1. But I have yet to debug this and am not certain it is a regression. I am also skeptical about the great results for scan. I suspect that I have more work to do to make the benchmark less subject to variance from MVCC GC (vacuum here). I also struggle with that on RocksDB (compaction), but not on InnoDB (purge).

Relative to: Postgres 17.6

col-1 : 18beta3

col-2 : 18rc1

col-1   col-2   point queries

1.01    0.98    hot-points_range=100

1.01    1.00    point-query_range=100

1.02    1.02    points-covered-pk_range=100

0.99    1.01    points-covered-si_range=100

1.00    0.99    points-notcovered-pk_range=100

1.00    0.99    points-notcovered-si_range=100

1.01    1.00    random-points_range=1000

1.01    0.99    random-points_range=100

1.01    1.00    random-points_range=10

col-1   col-2   range queries without aggregation

0.97    0.96    range-covered-pk_range=100

0.97    0.97    range-covered-si_range=100

0.99    0.99    range-notcovered-pk_range=100

0.99    0.99    range-notcovered-si_range=100

1.35    1.36    scan_range=100

col-1   col-2   range queries with aggregation

1.02    1.03    read-only-count_range=1000

1.00    1.00    read-only-distinct_range=1000

0.99    0.99    read-only-order_range=1000

1.00    1.00    read-only_range=10000

1.00    0.99    read-only_range=100

0.99    0.98    read-only_range=10

1.01    1.01    read-only-simple_range=1000

1.02    1.00    read-only-sum_range=1000

col-1   col-2   writes

0.99    0.99    delete_range=100

0.99    1.01    insert_range=100

0.99    0.99    read-write_range=100

0.99    0.99    read-write_range=10

0.98    0.98    update-index_range=100

1.00    0.99    update-inlist_range=100

0.98    0.98    update-nonindex_range=100

0.98    0.97    update-one_range=100

0.98    0.97    update-zipf_range=100

0.99    0.98    write-only_range=10000

Results: large32 server

I don't see small regressions in throughput for range queries without aggregation across Postgres 18 beta1, beta2, beta3 and rc1. I have only seen that on the low concurrency (small server) results.

The improvements on the scan microbenchmark come from using less CPU. But I am skeptical about the improvements. I might have more work to do to make the benchmark less subject to variance from MVCC GC (vacuum here). I also struggle with that on RocksDB (compaction), but not on InnoDB (purge).

Relative to: Postgres 17.6

col-1 : Postgres 18rc1

col-1   point queries

1.01    hot-points_range=100

1.01    point-query_range=100

1.01    points-covered-pk_range=100

1.01    points-covered-si_range=100

1.00    points-notcovered-pk_range=100

1.00    points-notcovered-si_range=100

1.01    random-points_range=1000

1.00    random-points_range=100

1.01    random-points_range=10

col-1   range queries without aggregation

0.99    range-covered-pk_range=100

0.99    range-covered-si_range=100

0.99    range-notcovered-pk_range=100

0.99    range-notcovered-si_range=100

1.12    scan_range=100

col-1   range queries with aggregation

1.00    read-only-count_range=1000

1.02    read-only-distinct_range=1000

1.01    read-only-order_range=1000

1.03    read-only_range=10000

1.00    read-only_range=100

1.00    read-only_range=10

1.00    read-only-simple_range=1000

1.00    read-only-sum_range=1000

col-1   writes

1.01    delete_range=100

1.00    insert_range=100

1.00    read-write_range=100

1.00    read-write_range=10

1.00    update-index_range=100

1.00    update-inlist_range=100

1.00    update-nonindex_range=100

0.99    update-one_range=100

1.00    update-zipf_range=100

1.00    write-only_range=10000

Results: large48 server

I don't see small regressions in throughput for range queries without aggregation across Postgres 18 beta1, beta2, beta3 and rc1. I have only seen that on the low concurrency (small server) results.

The improvements on the scan microbenchmark come from using less CPU. But I am skeptical about the improvements. I might have more work to do to make the benchmark less subject to variance from MVCC GC (vacuum here). I also struggle with that on RocksDB (compaction), but not on InnoDB (purge).

I am skeptical about the regression I see here for scan. That comes from using ~10% more CPU per query. I might have more work to do to make the benchmark less subject to variance from MVCC GC (vacuum here). I also struggle with that on RocksDB (compaction), but not on InnoDB (purge).

I have not see the large improvements for the insert and delete microbenchmarks on previous tests on that large server. I assume this is another case where I need to figure out how to reduce variance when I run the benchmark.

Relative to: Postgres 17.6

col-1 : Postgres 18beta3

col-2 : Postgres 18rc1

col-1   col-2   point queries

0.99    0.99    hot-points_range=100

0.99    0.99    point-query_range=100

1.00    0.99    points-covered-pk_range=100

0.99    1.02    points-covered-si_range=100

1.00    0.99    points-notcovered-pk_range=100

0.99    1.01    points-notcovered-si_range=100

1.00    0.99    random-points_range=1000

1.00    0.99    random-points_range=100

1.00    1.00    random-points_range=10

col-1   col-2   range queries without aggregation

0.99    0.99    range-covered-pk_range=100

0.98    0.99    range-covered-si_range=100

0.99    0.99    range-notcovered-pk_range=100

1.01    1.01    range-notcovered-si_range=100

0.91    0.91    scan_range=100

col-1   col-2   range queries with aggregation

1.04    1.03    read-only-count_range=1000

1.02    1.01    read-only-distinct_range=1000

1.01    1.00    read-only-order_range=1000

1.06    1.06    read-only_range=10000

0.98    0.97    read-only_range=100

0.99    0.99    read-only_range=10

1.02    1.02    read-only-simple_range=1000

1.03    1.03    read-only-sum_range=1000

col-1   col-2   writes

1.46    1.49    delete_range=100

1.32    1.32    insert_range=100

0.99    1.00    read-write_range=100

0.98    1.00    read-write_range=10

0.99    1.00    update-index_range=100

0.95    1.03    update-inlist_range=100

1.00    1.02    update-nonindex_range=100

0.96    1.04    update-one_range=100

1.00    1.01    update-zipf_range=100

1.00    1.00    write-only_range=10000

Postgres 18 beta3, large server, sysbench

This has performance results for Postgres 18 beta3, beta2, beta1, 17.5 and 17.4 using the sysbench benchmark and a large server. The working set is cached and the benchmark is run with high concurrency (40 connections). The goal is to search for CPU and mutex regressions. This work was done by Small Datum LLC and not sponsored

tl;dr

• There might be small regressions (~2%) for several range queries that don't do aggregation. This is similar to what I reported for 18 beta3 on a small server, but here it only occurs for 3 of the 4 microbenchmarks and on the small server it occurs on all 4. I am still uncertain about whether this really is a regression.

Builds, configuration and hardware

I compiled Postgres versions 17.4, 17.5, 18 beta1, 18 beta2 and 18 beta3 from source.

The server is an ax162-s from Hetzner with an AMD EPYC 9454P processor, 48 cores, AMD SMT disabled and 128G RAM. The OS is Ubuntu 22.04. Storage is 2 NVMe devices with SW RAID 1 and 

ext4. More details on it are here.

The config file for Postgres 17.4 and 17.5 is x10a_c32r128.

The config files for Postgres 18 are:

• x10b_c32r128 is functionally the same as x10a_c32r128 but adds io_method=sync
• x10d_c32r128 starts with x10a_c2r128 and adds io_method=io_uring

Benchmark

I used sysbench and my usage is explained here. To save time I only run 32 of the 42 microbenchmarks and most test only 1 type of SQL statement. Benchmarks are run with the database cached by Postgres.

The tests are run using 8 tables with 10M rows per table. The read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

Results

The microbenchmarks are split into 4 groups -- 1 for point queries, 2 for range queries, 1 for writes. For the range query microbenchmarks, part 1 has queries that don't do aggregation while part 2 has queries that do aggregation. 

I provide charts below with relative QPS. The relative QPS is the following:

(QPS for some version) / (QPS for Postgres 17.5)

When the relative QPS is > 1 then some version is faster than PG 17.5.  When it is < 1 then there might be a regression. Values from iostat and vmstat divided by QPS are also provided here. These can help to explain why something is faster or slower because it shows how much HW is used per request.

Relative to: pg174_o2nofp.x10a_c32r128

col-1 : pg175_o2nofp.x10a_c32r128

col-2 : pg18beta1_o2nofp.x10b_c32r128

col-3 : pg18beta1_o2nofp.x10d_c32r128

col-4 : pg18beta2_o2nofp.x10d_c32r128

col-5 : pg18beta3_o2nofp.x10d_c32r128

col-1   col-2   col-3   col-4   col-5

0.98    0.99    0.99    1.00    0.99    hot-points_range=100

1.01    1.01    1.00    1.01    1.01    point-query_range=100

1.00    1.00    0.99    1.00    1.00    points-covered-pk

1.00    1.01    1.00    1.02    1.00    points-covered-si

1.00    1.01    1.00    1.00    1.00    points-notcovered-pk

1.00    1.00    1.01    1.02    1.00    points-notcovered-si

1.00    1.00    1.00    1.00    1.00    random-points_range=1000

1.00    1.01    1.00    1.00    1.00    random-points_range=100

1.00    1.00    1.00    1.00    1.00    random-points_range=10

1.00    0.97    0.96    0.98    0.97    range-covered-pk

1.00    0.97    0.97    0.98    0.97    range-covered-si

0.99    0.99    0.99    0.99    0.98    range-notcovered-pk

1.00    1.01    1.01    1.00    1.01    range-notcovered-si

1.00    1.02    1.03    1.03    1.02    read-only-count

1.00    1.00    1.00    1.01    1.01    read-only-distinct

1.00    1.00    1.00    1.00    1.00    read-only-order

1.01    1.01    1.02    1.02    1.01    read-only_range=10000

1.00    0.99    0.99    0.99    1.00    read-only_range=100

1.01    0.99    0.99    1.00    0.99    read-only_range=10

1.00    1.01    1.01    1.01    1.01    read-only-simple

1.00    1.02    1.03    1.03    1.02    read-only-sum

1.00    1.13    1.14    1.02    0.91    scan_range=100

1.00    1.13    1.13    1.02    0.90    scan.warm_range=100

1.00    0.99    0.99    0.99    0.99    delete_range=100

0.99    1.00    1.02    0.99    1.00    insert_range=100

1.01    1.00    1.00    1.00    0.99    read-write_range=100

1.00    0.98    1.00    1.01    0.99    read-write_range=10

0.99    0.99    1.02    0.98    0.96    update-index

1.00    1.01    1.00    1.00    1.01    update-inlist

0.98    0.98    0.99    0.98    0.97    update-nonindex

0.95    0.95    0.94    0.93    0.95    update-one_range=100

0.97    0.98    0.98    0.97    0.95    update-zipf_range=100

0.98    0.99    0.99    0.98    0.98    write-only_range=10000

Postgres 18 beta3, small server, sysbench

This has performance results for Postgres 18 beta3, beta2, beta1 and 17.6 using the sysbench benchmark and a small server. The working set is cached and the benchmark is run with low concurrency (1 connection). The goal is to search for CPU regressions. This work was done by Small Datum LLC and not sponsored

tl;dr

• There might be small regressions (~2%) for several range queries that don't do aggregation. This is similar to what I reported for 18 beta1.
• Vacuum continues to be a problem for me and I had to repeat the benchmark a few times to get a stable result. It appears to be a big source of non-deterministic behavior leading to false alarms for CPU regressions in read-heavy tests that run after vacuum. In some ways, RocksDB compaction causes similar problems. Fortunately, InnoDB MVCC GC (purge) does not cause such problems.

Builds, configuration and hardware

I compiled Postgres versions 17.6, 18 beta1, 18 beta2 and 18 beta3 from source.

The server is a Beelink SER7 with a Ryzen 7 7840HS CPU, 32G of RAM, 8 cores with AMD SMT disabled, Ubuntu 24.04 and an NVMe devices with discard enabled and ext4 for the database.

The config file for Postgres 17.6 is x10a_c8r32.

The config files for Postgres 18 are:

• x10b_c8r32 is functionally the same as x10a_c8r32 but adds io_method=sync
• x10b1_c8r32 starts with x10b_c8r32 and adds vacuum_max_eager_freeze_failure_rate =0
• x10b2_c8r32 starts with x10b_c8r32 and adds vacuum_max_eager_freeze_failure_rate =0.99

Benchmark

I used sysbench and my usage is explained here. To save time I only run 32 of the 42 microbenchmarks and most test only 1 type of SQL statement. Benchmarks are run with the database cached by Postgres.

The tests are run using 1 table with 50M rows. The read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

Results

The microbenchmarks are split into 4 groups -- 1 for point queries, 2 for range queries, 1 for writes. For the range query microbenchmarks, part 1 has queries that don't do aggregation while part 2 has queries that do aggregation. 

I provide charts below with relative QPS. The relative QPS is the following:

(QPS for some version) / (QPS for Postgres 17.6)

When the relative QPS is > 1 then some version is faster than PG 17.6.  When it is < 1 then there might be a regression. Values from iostat and vmstat divided by QPS are also provided here. These can help to explain why something is faster or slower because it shows how much HW is used per request.

The numbers highlighted in yellow below might be from a small regression for range queries that don't do aggregation. But note that this does reproduce for the full table scan microbenchmark (scan). I am not certain it is a regression as this might be from non-deterministic CPU overheads for read-heavy workloads that are run after vacuum. I hope to look at CPU flamegraphs soon.

• the mapping from microbenchmark name to Lua script is here
• the range query without aggregation microbenchmarks use oltp_range_covered.lua with various flags set and the SQL statements it uses are here. All of these return 100 rows.
• the scan microbenchmark uses oltp_scan.lua which is a SELECT with a WHERE clause that filters all rows (empty result set)

Relative to: x.pg176_o2nofp.x10a_c8r32.pk1

col-1 : x.pg18beta1_o2nofp.x10b_c8r32.pk1

col-2 : x.pg18beta2_o2nofp.x10b_c8r32.pk1

col-3 : x.pg18beta3_o2nofp.x10b_c8r32.pk1

col-4 : x.pg18beta3_o2nofp.x10b1_c8r32.pk1

col-5 : x.pg18beta3_o2nofp.x10b2_c8r32.pk1

col-1   col-2   col-3   col-4   col-5 -> point queries

1.00    1.00    0.98    0.99    0.99    hot-points_range=100

1.00    1.01    1.00    1.00    0.99    point-query_range=100

1.00    1.02    1.01    1.01    1.01    points-covered-pk

1.00    1.00    1.00    1.00    1.00    points-covered-si

1.01    1.01    1.00    1.00    1.00    points-notcovered-pk

1.01    1.00    1.00    1.00    1.00    points-notcovered-si

0.99    1.00    0.99    1.00    1.00    random-points_range=1000

1.01    1.00    1.00    1.00    1.00    random-points_range=100

1.01    1.01    1.00    1.00    0.99    random-points_range=10

col-1   col-2   col-3   col-4   col-5 -> range queries w/o agg

0.98    0.99    0.97    0.98    0.96    range-covered-pk_range=100

0.98    0.99    0.96    0.98    0.97    range-covered-si_range=100

0.98    0.98    0.98    0.97    0.98    range-notcovered-pk

0.99    0.99    0.98    0.98    0.98    range-notcovered-si

1.01    1.02    1.00    1.00    1.00    scan

col-1   col-2   col-3   col-4   col-5 -> range queries with agg

1.02    1.01    1.02    1.01    0.98    read-only-count_range=1000

0.98    1.01    1.01    1.00    1.03    read-only-distinct

0.99    0.99    0.99    0.99    0.99    read-only-order_range=1000

1.00    1.00    1.01    1.00    1.01    read-only_range=10000

0.99    0.99    0.99    0.99    0.99    read-only_range=100

0.99    0.99    0.99    0.98    0.99    read-only_range=10

1.01    1.00    1.00    1.00    1.01    read-only-simple

1.01    1.00    1.01    1.00    1.00    read-only-sum_range=1000

col-1   col-2   col-3   col-4   col-5 -> writes

0.99    1.00    0.98    0.98    0.98    delete_range=100

0.99    0.98    0.98    1.00    0.98    insert_range=100

0.99    0.99    0.99    0.98    0.99    read-write_range=100

0.98    0.99    0.99    0.98    0.99    read-write_range=10

1.00    0.99    0.98    0.97    0.99    update-index_range=100

1.01    1.00    0.99    1.01    1.00    update-inlist_range=100

1.00    1.00    0.99    0.96    0.99    update-nonindex_range=100

1.01    1.01    0.99    0.97    0.99    update-one_range=100

1.00    1.00    0.99    0.98    0.99    update-zipf_range=100

1.00    0.99    0.98    0.98    1.00    write-only_range=10000

MySQL 5.6 thru 9.4: small server, Insert Benchmark

This has results for the Insert Benchmark on a small server with InnoDB from MySQL 5.6 through 9.4. The workload here uses low concurrency (1 client), a small server and a cached database. I run it this way to look for CPU regressions before moving on to IO-bound workloads with high concurrency.

tl;dr

• good news - there are no large regressions after MySQL 8.0
• bad news - there are large regressions from MySQL 5.6 to 5.7 to 8.0
• load in 8.0, 8.4 and 9.4 gets about 60% of the throughput vs 5.6
• queries in 8.0, 8.4 and 9.4 get between 60% and 70% of the throughput vs 5.6

Builds, configuration and hardware

I compiled MySQL 5.6.51, 5.7.44, 8.0.43, 8.4.6 and 9.4.0 from source.

The server is an ASUS PN53 with 8 cores, AMD SMT disabled and 32G of RAM. The OS is Ubuntu 24.04. Storage is 1 NVMe device with ext4. More details on it are here.

I used the cz12a_c8r32 config file (my.cnf) which is here for 5.6.51, 5.7.44, 8.0.43, 8.4.6 and 9.4.0.

The Benchmark

The benchmark is explained here. I recently updated the benchmark client to connect via socket rather than TCP so that I can get non-SSL connections for all versions tested. AFAIK, with TCP I can only get SSL connections for MySQL 8.4 and 9.4.

The workload uses 1 client, 1 table with 30M rows and a cached database.

The benchmark steps are:

• l.i0
• insert 30 million rows per table in PK order. The table has a PK index but no secondary indexes. There is one connection per client.
• l.x
• create 3 secondary indexes per table. There is one connection per client.
• l.i1
• use 2 connections/client. One inserts 40 million rows per table and the other does deletes at the same rate as the inserts. Each transaction modifies 50 rows (big transactions). This step is run for a fixed number of inserts, so the run time varies depending on the insert rate.
• l.i2
• like l.i1 but each transaction modifies 5 rows (small transactions) and 10 million rows are inserted and deleted per table.
• Wait for N seconds after the step finishes to reduce variance during the read-write benchmark steps that follow. The value of N is a function of the table size.
• qr100
• use 3 connections/client. One does range queries and performance is reported for this. The second does does 100 inserts/s and the third does 100 deletes/s. The second and third are less busy than the first. The range queries use covering secondary indexes. This step is run for 1800 seconds. If the target insert rate is not sustained then that is considered to be an SLA failure. If the target insert rate is sustained then the step does the same number of inserts for all systems tested.
• qp100
• like qr100 except uses point queries on the PK index
• qr500
• like qr100 but the insert and delete rates are increased from 100/s to 500/s
• qp500
• like qp100 but the insert and delete rates are increased from 100/s to 500/s
• qr1000
• like qr100 but the insert and delete rates are increased from 100/s to 1000/s
• qp1000
• like qp100 but the insert and delete rates are increased from 100/s to 1000/s

Results: overview

The performance report is here.

The summary section has 3 tables. The first shows absolute throughput by DBMS tested X benchmark step. The second has throughput relative to the version from the first row of the table. The third shows the background insert rate for benchmark steps with background inserts. The second table makes it easy to see how performance changes over time. The third table makes it easy to see which DBMS+configs failed to meet the SLA. The summary section is here.

Below I use relative QPS (rQPS) to explain how performance changes. It is: (QPS for $me / QPS for $base) where $me is the result for some version $base is the result from MySQL 5.6.51.

When rQPS is > 1.0 then performance improved over time. When it is < 1.0 then there are regressions. When it is 0.90 then I claim there is a 10% regression. The Q in relative QPS measures: 

• insert/s for l.i0, l.i1, l.i2
• indexed rows/s for l.x
• range queries/s for qr100, qr500, qr1000
• point queries/s for qp100, qp500, qp1000

Below I use colors to highlight the relative QPS values with yellow for regressions and blue for improvements.

Results: details

This table is a copy of the second table in the summary section. It lists the relative QPS (rQPS) for each benchmark step where rQPS is explained above.

The benchmark steps are explained above, they are:

• l.i0 - initial load in PK order
• l.x - create 3 secondary indexes per table
• l.i1, l.i2 - random inserts and random deletes
• qr100, qr500, qr1000 - short range queries with background writes
• qp100, qp500, qp1000 - point queries with background writes

dbms	l.i0	l.x	l.i1	l.i2	qr100	qp100	qr500	qp500	qr1000	qp1000
5.6.51	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00	1.00
5.7.44	0.89	1.52	1.14	1.08	0.83	0.84	0.83	0.84	0.84	0.84
8.0.43	0.60	2.50	1.04	0.86	0.69	0.62	0.69	0.63	0.70	0.62
8.4.6	0.60	2.53	1.03	0.86	0.68	0.61	0.67	0.61	0.68	0.61
9.4.0	0.60	2.53	1.03	0.87	0.70	0.63	0.70	0.63	0.70	0.62

The summary is:

• l.i0
• there are large regressions starting in 8.0 and modern MySQL only gets ~60% of the throughput relative to 5.6 because modern MySQL has more CPU overhead
• l.x
• I ignore this but there have been improvements
• l.i1, l.i2
• there was a large improvement in 5.7 but new CPU overhead since 8.0 reduces that
• qr100, qr500, qr1000
• there are large regressions from 5.6 to 5.7 and then again from 5.7 to 8.0
• throughput in modern MySQL is ~60% to 70% of what it was in 5.6

Sysbench for MySQL 5.6 thru 9.4 on a small server

This has performance results for InnoDB from MySQL 5.6.51, 5.7.44, 8.0.43, 8.4.6 and 9.4.0 on a small server with sysbench microbenchmarks. The workload here is cached by InnoDB and my focus is on regressions from new CPU overheads. This work was done by Small Datum LLC and not sponsored. 

tl;dr

• Low concurrency (1 client) is the worst case for regressions in modern MySQL
• MySQL 8.0, 8.4 and 9.4 are much slower than 5.6.51 in all but 2 of the 32 microbenchmarks
• The bad news - performance regressions aren't getting fixed
• The good news - regressions after MySQL 8.0 are small

Builds, configuration and hardware

I compiled MySQL from source for versions 5.6.51, 5.7.44, 8.0.43, 8.4.6 and 9.4.0.

The server is an ASUS ExpertCenter PN53 with AMD Ryzen 7 7735HS, 32G RAM and an m.2 device for the database. More details on it are here. The OS is Ubuntu 24.04 and the database filesystem is ext4 with discard enabled.

The my.cnf files are here.

Benchmark

I used sysbench and my usage is explained here. To save time I only run 32 of the 42 microbenchmarks and most test only 1 type of SQL statement. Benchmarks are run with the database cached by InnoDB.

The tests are run using 1 table with 50M rows. The read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

Results

All files I saved from the benchmark are here and the spreadsheet is here.

The microbenchmarks are split into 4 groups -- 1 for point queries, 2 for range queries, 1 for writes. For the range query microbenchmarks, part 1 has queries that don't do aggregation while part 2 has queries that do aggregation. 

I provide charts below with relative QPS. The relative QPS is the following:

(QPS for some version) / (QPS for MySQL 5.6.51)

When the relative QPS is > 1 then some version is faster than 5.6.51.  When it is < 1 then there might be a regression. Values from iostat and vmstat divided by QPS are also provided here. These can help to explain why something is faster or slower because it shows how much HW is used per request.

Results: point queries

Based on results from vmstat the regressions are from new CPU overheads.

Results: range queries without aggregation

Based on results from vmstat the regressions are from new CPU overheads.

Results; range queries with aggregation

Based on results from vmstat the regressions are from new CPU overheads.

Results: writes

Based on results from vmstat the regressions are from new CPU overheads.

Postgres 18 beta2: large server, Insert Benchmark, part 2

I repeated the benchmark for one of the workloads used in a recent blog post on Postgres 18 beta2 performance. The workload used 1 client and 1 table with 50M rows that fits in the Postgres buffer pool. In the result I explain here, one of the benchmark steps was run for ~10X more time. Figuring out how long to run the steps in the Insert Benchmark is always a work in progress -- I want to test more things, so I don't want to run steps for too long, but there will be odd results if the run times are too short.

tl;dr

• up to 2% less throughput on range queries in the qr100 benchmark step. This is similar to what I saw in my previous report.
• up to 12% more throughput on the l.i2 benchmark step in PG beta1 and beta2. This is much better than what I saw in my previous report.

Details

Details on the benchmark are in my previous post.

The benchmark is explained here and was run for one workloads -- 1 client, cached.

• run with 1 client, 1 table and a cached database
• load 50M rows in step l.i0, do 160M writes in step l.i1 and 40M in l.i2. Note that here the l.i1 and l.i2 steps run for ~10X longer than in my previous post.

The benchmark steps are:

• l.i0
• insert X million rows per table in PK order. The table has a PK index but no secondary indexes. There is one connection per client.
• l.x
• create 3 secondary indexes per table. There is one connection per client.
• l.i1
• use 2 connections/client. One inserts Y million rows per table and the other does deletes at the same rate as the inserts. Each transaction modifies 50 rows (big transactions). This step is run for a fixed number of inserts, so the run time varies depending on the insert rate.
• l.i2
• like l.i1 but each transaction modifies 5 rows (small transactions) and Z million rows are inserted and deleted per table.
• Wait for N seconds after the step finishes to reduce variance during the read-write benchmark steps that follow. The value of N is a function of the table size.
• qr100
• use 3 connections/client. One does range queries and performance is reported for this. The second does does 100 inserts/s and the third does 100 deletes/s. The second and third are less busy than the first. The range queries use covering secondary indexes. This step is run for 1800 seconds. If the target insert rate is not sustained then that is considered to be an SLA failure. If the target insert rate is sustained then the step does the same number of inserts for all systems tested.
• qp100
• like qr100 except uses point queries on the PK index
• qr500
• like qr100 but the insert and delete rates are increased from 100/s to 500/s
• qp500
• like qp100 but the insert and delete rates are increased from 100/s to 500/s
• qr1000
• like qr100 but the insert and delete rates are increased from 100/s to 1000/s
• qp1000
• like qp100 but the insert and delete rates are increased from 100/s to 1000/s

Results: overview

The performance report is here.

The summary section has 3 tables. The first shows absolute throughput by DBMS tested X benchmark step. The second has throughput relative to the version from the first row of the table. The third shows the background insert rate for benchmark steps with background inserts. The second table makes it easy to see how performance changes over time. The third table makes it easy to see which DBMS+configs failed to meet the SLA.

Below I use relative QPS (rQPS) to explain how performance changes. It is: (QPS for $me / QPS for $base) where $me is the result for some version $base is the result from Postgres 17.4.

When rQPS is > 1.0 then performance improved over time. When it is < 1.0 then there are regressions. When it is 0.90 then I claim there is a 10% regression. The Q in relative QPS measures: 

• insert/s for l.i0, l.i1, l.i2
• indexed rows/s for l.x
• range queries/s for qr100, qr500, qr1000
• point queries/s for qp100, qp500, qp1000

Below I use colors to highlight the relative QPS values with red for <= 0.97, green for >= 1.03 and grey for values between 0.98 and 1.02.

Results: 1 client, cached

Normally I summarize the summary but I don't do that here to save space.

But the tl;dr is:

• up to 2% less throughput on range queries in the qr100 benchmark step. This is similar to what I saw in my previous report.
• up to 12% more throughput on the l.i2 benchmark step in PG beta1 and beta2. This is much better than what I saw in my previous report.

Postgres 18 beta2: large server, sysbench

This has performance results for Postgres 17.4, 17.5, 18 beta1 and 18 beta2 on a large server with sysbench microbenchmarks. Results like this from me are usually boring because Postgres has done a great job at avoiding performance regressions over time. This work was done by Small Datum LLC and not sponsored. Previous work from me for Postgres 17.4 and 18 beta1 is here.

The workload here is cached by Postgres and my focus is on regressions from new CPU overhead or mutex contention.

tl;dr

• there might be small regressions (~2%) for range queries on the benchmark with 1 client. One cause is more CPU in BuildCachedPlan. 
• there might besmall regressions (~2%) for range queries on the benchmark with 40 clients. One cause is more CPU in PortalRunSelect.
• otherwise things look great

Builds, configuration and hardware

I compiled Postgres from source using -O2 -fno-omit-frame-pointer.

The server is an ax162-s from Hetzner with an AMD EPYC 9454P processor, 48 cores, AMD SMT disabled and 128G RAM. The OS is Ubuntu 22.04. Storage is 2 NVMe devices with SW RAID 1 and 

ext4. More details on it are here.

The config file for 17.4 and 17.5 is conf.diff.cx10a_c32r128.

The config files for 18 beta 1 are:

• conf.diff.cx10b_c8r32
• uses io_method='sync' to match Postgres 17 behavior
• conf.diff.cx10c_c8r32
• uses io_method='worker' and io_workers=32 to do async IO via a thread pool. I eventually learned that 32 is too large but I don't think it matters much on this workload.
• conf.diff.cx10d_c8r32
• uses io_method='io_uring' to do async IO via io_uring

Benchmark

I used sysbench and my usage is explained here. To save time I only run 27 of the 42 microbenchmarks and most test only 1 type of SQL statement. Benchmarks are run with the database cached by Postgres.

The tests are run using two workloads. For both the read-heavy microbenchmarks run for 300 seconds and write-heavy run for 600 seconds.

• 1-client
• run with 1 client and 1 table with 50M rows
• 40-clients
• run with 40 client and 8 table with 10M rows per table

The command lines to run all tests with my helper scripts are:

•  bash r.sh 1 50000000 300 600 $deviceName 1 1 1
• bash r.sh 8 10000000 300 600 $deviceName 1 1 40

Results

All files I saved from the benchmark are here.

I don't provide graphs in this post to save time and because there are few to no regressions from Postgres 17.4 to 18 beta2. The microbenchmarks are split into 4 groups -- 1 for point queries, 2 for range queries, 1 for writes. For the range query microbenchmarks, part 1 has queries that don't do aggregation while part 2 has queries that do aggregation. 

I provide tables below with relative QPS. The relative QPS is the following:

(QPS for some version) / (QPS for PG 17.4)

When the relative QPS is > 1 then some version is faster than 17.4.  When it is < 1 then there might be a regression. Values from iostat and vmstat divided by QPS are also provided. These can help to explain why something is faster or slower because it shows how much HW is used per request.

Results: 1-client

Tables with QPS per microbenchmark are here using absolute and relative QPS. All of the files I saved for this workload are here.

For point queries

• QPS is mostly ~2% better in PG 18 beta2 relative to 17.4 (see here), but the same is true for 17.5. Regardless, this is good news.

For range queries without aggregation

• full table scan is ~6% faster in PG 18 beta2 and ~4% faster in 17.5, both relative to 17.4
• but for the other microbenchmarks, PG 18 beta2, 18 beta1 and 17.5 are 1% to 5% slower than 17.4. 
• From vmstat and iostat metrics for range-[not]covered-pk and range-[not]covered-si this is explained by an increase in CPU/query (see the cpu/o column in the previous links). I also see a few cases where CPU/query is much larger but only for 18 beta2 with configs that use io_method =worker and =io_uring. 
• I measured CPU using vmstat which includes all CPU on the host so perhaps something odd happens with other Postgres processes or some rogue process is burning CPU. I checked more results from vmstat and iostat and don't see storage IO during the tests.
• Code that does the vacuum and checkpoint is here, output from the vacuum work is here, and the Postgres logfiles are here. This work is done prior to the range query tests.

For range queries with aggregation

• there are regressions (see here), but here they are smaller than what I see above for range queries without aggregation
• the interesting result is for the same query, but run with different selectivity to go from a larger to a smaller range and the regression increases as the range gets smaller (see here). To me this implies the likely problem is the fixed cost -- either in the optimizer or query setup (allocating memory, etc).

For writes

• there are small regressions, mostly from 1% to 3% (see here).
• the regressions are largest for the 18beta configs that use io_method=io_uring, that might be expected given the benefits it provides

Then I used Flamegraphs (see here) to try and explain the regressions. My benchmark helper scripts collect a Flamegraph about once per minute for each microbenchmark and the microbenchmarks were run for 5 minutes if read-mostly or 10 minutes if write-heavy. Then the ~5 or ~10 samples from perf (per microbenchmark) are combined to produce one Flamegraph per microbenchmark. My focus is on the distribution of time across thread stacks where there are stacks for parse, optimize, execute and network.

• For range-covered-pk there is a small (2% to 3%) increase from PG 17.4 to 18 beta2 in BuildCachedPlan (see here for 17.4 and 18 beta2).
• The increase in CPU for BuildCachedPlan also appears in Flamegraphs for other range query microbenchmarks

Results: 40-clients

Tables with QPS per microbenchmark are here using absolute and relative QPS. All of the files I saved for this workload are here, Postgres logfiles are here, output from vacuum is here and Flamegraphs are here.

For point queries

• QPS is similar from PG 17.4 through 18 beta1 (see here).

For range queries without aggregation

• full table scan is mostly ~2% faster after 17.4 (see here)
• for the other microbenchmarks, 3 of the 4 have small regressions of ~2% (see here). The worst is range-covered-pk and the problem appears to be more CPU per query (see here). Unlike above where the new overhead was in BuildCachedPlan, here it is in the stack with PortalRunSelect.

For range queries with aggregation

• QPS is similar from PG 17.4 through 18 beta2 (see here)

For writes

• QPS drops by 1% to 5% for many microbenchmarks, but this problem starts in 17.5 (see here)
• From vmstat and iostat metrics for update-one (which suffers the most, see here) the CPU per operation overhead does not increase (see the cpu/o column), the number of context switches per operation also does not increase (see the cs/o column).
• Also from iostat, the amount of data written to storage doesn't change much.