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 do 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 also provided. Theses 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.

Postgres 18 beta2: large server, Insert Benchmark

This has results for the Insert Benchmark with Postgres on a large server. 

There might be small regressions, but I have more work in progress to explain that:

• for a workload with 1 client and a cached database I see a small increase in CPU/operation (~10%) during the l.i2 benchmark step. I am repeating that benchmark.
• for a workload with 20 clients and an IO-bound database I see a small decrease in QPS (typically 2% to 4%) during read+write benchmark steps.

Builds, configuration and hardware

I compiled Postgres from source using -O2 -fno-omit-frame-pointer for versions 18 beta2, 18 beta1, 17.5 and 17.4.

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 here and named conf.diff.cx10a_c32r128.

For 18 beta1 and beta2 I tested 3 configuration files, and they are here:

• conf.diff.cx10b_c32r128 (x10b) - uses io_method=sync
• conf.diff.cx10cw4_c32r128 (x10cw4) - uses io_method=worker with io_workers=4
• conf.diff.cx10d_c32r128 (x10d) - uses io_method=io_uring

The Benchmark

The benchmark is explained here and was run for three workloads:

• 1 client, cached
• run with 1 client, 1 table and a cached database
• load 50M rows in step l.i0, do 16M writes in step l.i1 and 4M in l.i2
• 20 clients, cached
• run with 20 clients, 20 tables (table per client) and a cached database
• for each client/table - load 10M rows in step l.i0, do 16M writes in step l.i1 and 4M in l.i2
• 20 clients, IO-bound
• run with 20 clients, 20 tables (table per client) and a database larger than RAM
• for each client/table - load 200M rows in step l.i0, do 4M writes in step l.i1 and 1M in l.i2
• for the qr100, qr500 and qr1000 steps the working set is cached, otherwise it is not

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 reports are here:

• 1 client, cached
• 20 clients, cached
• 20 clients, IO-bound

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 sections are here:

• 1 client, cached
• 20 clients, cached
• 20 clients, IO-bound

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.

There might be regressions on the l.i2 benchmark step that does inserts+deletes with smaller transactions while l.i1 does the same but with larger transactions. These first arrive with Postgres 17.5 but I will ignore that because it sustained a higher rate on the preceding benchmark step (l.i1) and might suffer from more vacuum debt during l.i2.

From the response time table, 18 beta2 does better than 18 beta1 based on the 256us and 1ms columns.

From the vmstat and iostat metrics, there is a ~10% increase in CPU/operation starting in Postgres 17.5 -- the value in the cpupq column increases from 596 for PG 17.4 to ~660 starting in 17.5. With one client the l.i2 step finishes in ~500 seconds and that might be too short. I am repeating the bencchmark to run that step for 4X longer.

Results: 20 clients, cached

Normally I summarize the summary but I don't do that here to save space. Regardless, this is easy to summarize - there are small improvements (~4%) on the l.i1 and l.i2 benchmark steps and no regressions elsewhere.

Results: 20 clients, IO-bound

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

From the summary, Postgres did not sustain the target write rates during qp1000 and qr1000 but I won't claim that it should have been able to -- perhaps I need faster IO. The first table in the summary section uses a grey background to indicate that. Fortunately, all versions were able to sustain a similar write rate. This was a also a problem for some versions on the qp500 step.

For the l.i2 step there is an odd outlier with PG 18beta2 and the cx10cw4_c32r128 config (that uses io_method=worker). I will ignore that for now.

For many of the read+write tests (qp100, qr100, qp500, qr500, qp1000, qr1000) thoughput with PG 18 beta1 and beta2 is up to 5% less than for PG 17.4. The regression might be explained by a small increase in CPU/operation.