Why is RocksDB spending so much time handling page faults?

This week I was running benchmarks to understand how fast RocksDB could do IO, and then compared that to fio to understand the CPU overhead added by RocksDB. While looking at flamegraphs taken during the benchmark I was confused that about 20% of the samples were from page fault handling. This confused me at first.

The lesson here is to run your benchmark long enough to reach a steady state before you measure things or there will be confusion. And I was definitely confused when I first saw this. Perhaps my post saves time for the next person who spots this.

The workload is db_bench with a database size that is much larger than memory and read-only microbenchmarks for point lookups and range scans.

Then I wondered if this was a transient issue that occurs while RocksDB is warming up the block cache and growing process RSS until the block cache has been fully allocated.

While b-trees as used by Postgres and MySQL will do a large allocation at process start, RocksDB does an allocation per block read, and when the block is evicted then the allocation is free'd. This can be a stress test for a memory allocator which is why jemalloc and tcmalloc work better than glibc malloc for RocksDB. I revisit the mallocator topic every few years and my most recent post is here.

In this case I use RocksDB with jemalloc. Even though per-block allocations are transient, the memory used by jemalloc is mostly not transient. While there are cases where jemalloc an return memory to the OS, with my usage that is unlikely to happen.

Were I to let the benchmark run for a long enough time, then eventually jemalloc would finish getting memory from the OS. However, my tests were running for about 10 minutes and doing about 10,000 block reads per second while I had configured RocksDB to use a block cache that was at least 36G and the block size was 8kb. So my tests weren't running long enough for the block cache to fill, which means that during the measurement period:

• jemalloc was still asking for memory
• block cache eviction wasn't needed and after each block read a new entry was added to the block cache

The result in this example is 22.69% of the samples are from page fault handling. That is the second large stack from the left. The RocksDB code where it happens is rocksdb::BlockFetcher::ReadBlockContents.

When I run the benchmark for more time, the CPU overhead from page fault handling goes away.

Is it time for TPC-BLOB?

If you want to store vectors in your database then what you store as a row, KV pair or document is likely to be larger than the fixed-page size (when your DBMS uses fixed-page sizes) and you will soon care about efficient and performant support for large objects. I assume this support hasn't been the top priority for many DBMS implementations and there will be some performance bugs.

In a SQL DBMS, support for large objects will use the plumbing created to handle LOB (Large OBject) datatypes. We should define what the L in LOB means here and I will wave my hands and claim larger than a fixed-page in your favorite DBMS but smaller than 512kb because I limit my focus to online workloads.

Perhaps now is the time for industry standard benchmarks for workloads with large objects. Should it be TPC-LOB or TPC-BLOB?

Most popular DBMS use fixed-size pages whether that storage is index-organized via an update-in-place b-tree (InnoDB) or heap-organized (Postgres, Oracle). For rows that are larger than the page size, which is usually between 4kb and 16kb, the entire row or largest columns will be stored out of line and likely split across several pages in the out of line storage. When the row is read, additional reads will be done to gather all of the too-large parts from the out of line locations.

This approach is far from optimal as there will be more CPU overhead, more random IO and might be more wasted space. But this was good enough because support for LOBs wasn't a priority for these DBMS as their focus was on OLTP where rows were likely to be smaller than a fixed-size page.

Perhaps by luck, perhaps it was fate, but WiredTiger is a great fit for MongoDB because it is more flexible about page sizes. And it is more flexible because it isn't an update-in-place b-tree, instead it is a copy-on-write random (CoW-R) b-tree that doesn't need or use out-of-line storage, although for extra large documents there might be a benefit from out-of-line.

MyRocks, and other LSM-based DBMS, also don't require out-of-line storage but they can benefit from it as shown by WiscKey and other engines that do key-value separation. Even the mighty RocksDB has an implementation of key-value separation via BlobDB.

Postgres 18.0 vs sysbench on a 32-core server

This is yet another great result for Postgres 18.0 vs sysbench. This time I used a 32-core server. Results for a 24-core server are here. The goal for this benchmark is to check for regressions from new CPU overhead and mutex contention.

I repeated the benchmark twice because I had some uncertainty about platform variance (HW and SW) on the first run.

tl;dr, from Postgres 17.6 to 18.0

• There might be regressions from 17.6 to 18.0 but they are small (usually <= 3%)

tl;dr, from Postgres 12.22 through 18.0

• the hot-points test is almost 2X faster starting in 17.6
• scan is ~1.2X faster starting in 14.19
• all write tests are much faster staring in 17.6

Builds, configuration and hardware

I compiled Postgres from source for versions 12.22, 13.22, 14.19, 15.14, 16.10, 17.6, and 18.0.

The server is a Dell Precision 7865 Tower Workstation with 1 socket, 128G RAM and an AMD Ryzen Threadripper PRO 5975WX with 32-Cores. The OS is Ubuntu 24.04 and storage is a 2TB m.2 SSD with ext-4 and discard enabled.

Prior to 18.0, the configuration file was named conf.diff.cx10a_c32r128 and is here for 12.22, 13.22, 14.19, 15.14, 16.10 and 17.6.

For 18.0 I tried 3 configuration files:

• conf.diff.cx10b_c32r128 (x10b) - uses io_method=sync
• conf.diff.cx10c_c32r128 (x10c) - uses io_method=worker
• conf.diff.cx10d_c32r128 (x10d) - uses 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 read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

The benchmark is run with 24 clients and 8 tables with 10M rows per table. The purpose is to search for regressions from new CPU overhead and mutex contention.

I ran the benchmark twice. In the first run, there was several weeks between getting results for the older Postgres releases and Postgres 18.0 so I am less certain about variance from the hardware and softare. One concern is changes in daily temperature because I don't have a climate-controlled server room. Another concern is changes from updating my OS install.

In the second run, all results were collected within 7 days and I am less concerned about variance there.

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 base version)

When the relative QPS is > 1 then some version is faster than base version.  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.

I present results for:

• versions 12 through 18 using 12.22 as the base version
• versions 17.6 and 18.0 using 17.6 as the base version

Results: Postgres 17.6 and 18.0

All files are here.

Results per microbenchmark from vmstat and iostat are here for the first and second run.

Some comments:

• 18.0 looks better relative to 17.6 in the second run and I explain my uncertainty about the first run above
• But I am skeptical about the great result for 18.0 on the full scan test (scan_range=100) in the second run. That might be variance induced by vacuum.
• There might be regressions from 17.6 to 18.0 but they are small (usually <= 3%)
• The small regression in read-only_range=10 might be from new optimizer overhead, because it doesn't reproduce when the length of the range query is increased -- see read-only_range=100 and read-only_range=10000.

Relative to: 17.6

col-1 : 18.0 with the x10b config that uses io_method=sync

col-2 : 18.0 with the x10c config that uses io_method=worker

col-3 : 18.0 with the x10d config that uses io_method=io_uring

col-1   col-2   col-3   point queries, first run

0.97    0.99    0.94    hot-points_range=100

0.97    0.98    0.96    point-query_range=100

1.00    0.99    0.99    points-covered-pk_range=100

0.99    1.00    1.00    points-covered-si_range=100

0.98    0.99    0.98    points-notcovered-pk_range=100

0.99    0.99    0.99    points-notcovered-si_range=100

1.00    1.00    0.99    random-points_range=1000

0.98    0.98    0.98    random-points_range=100

0.99    0.98    0.99    random-points_range=10

col-1   col-2   col-3   point queries, second run

0.98    1.00    0.99    hot-points_range=100

1.00    1.00    0.99    point-query_range=100

1.01    1.01    1.01    points-covered-pk_range=100

1.00    1.01    1.00    points-covered-si_range=100

1.00    0.98    1.00    points-notcovered-pk_range=100

1.00    1.00    1.01    points-notcovered-si_range=100

1.00    1.01    1.01    random-points_range=1000

1.00    0.99    1.01    random-points_range=100

0.99    0.99    1.00    random-points_range=10

col-1   col-2   col-3   range queries without aggregation, first run

0.97    0.98    0.95    range-covered-pk_range=100

0.97    0.97    0.94    range-covered-si_range=100

0.98    0.98    0.97    range-notcovered-pk_range=100

0.99    0.99    0.98    range-notcovered-si_range=100

0.97    0.99    0.96    scan_range=100

col-1   col-2   col-3   range queries without aggregation, second run

0.99    0.99    0.98    range-covered-pk_range=100

0.99    0.99    0.99    range-covered-si_range=100

0.98    0.99    0.98    range-notcovered-pk_range=100

0.99    1.00    1.00    range-notcovered-si_range=100

1.24    1.24    1.22    scan_range=100

col-1   col-2   col-3   range queries with aggregation, first run

0.99    1.00    1.00    read-only-count_range=1000

1.01    1.01    1.01    read-only-distinct_range=1000

1.01    1.01    1.00    read-only-order_range=1000

1.04    1.04    1.04    read-only_range=10000

0.99    0.99    0.98    read-only_range=100

0.97    0.98    0.97    read-only_range=10

0.99    0.98    0.98    read-only-simple_range=1000

0.99    0.99    0.99    read-only-sum_range=1000

col-1   col-2   col-3   range queries with aggregation, second run

0.99    1.00    1.00    read-only-count_range=1000

1.01    1.01    1.00    read-only-distinct_range=1000

0.99    0.99    1.00    read-only-order_range=1000

1.02    1.03    1.03    read-only_range=10000

0.99    0.99    0.99    read-only_range=100

0.99    0.99    0.98    read-only_range=10

0.99    1.00    1.01    read-only-simple_range=1000

1.00    1.00    1.00    read-only-sum_range=1000

col-1   col-2   col-3   writes, first run

0.99    0.98    0.96    delete_range=100

0.99    0.96    0.98    insert_range=100

1.00    0.99    0.98    read-write_range=100

0.99    0.98    0.98    read-write_range=10

1.00    0.99    1.00    update-index_range=100

1.03    0.95    1.01    update-inlist_range=100

0.99    0.99    1.00    update-nonindex_range=100

1.00    1.00    1.01    update-one_range=100

0.98    0.99    1.00    update-zipf_range=100

0.97    0.97    0.99    write-only_range=10000

col-1   col-2   col-3   writes, second run

0.97    0.97    0.98    delete_range=100

0.99    0.99    1.00    insert_range=100

0.99    0.99    0.98    read-write_range=100

0.98    0.98    0.98    read-write_range=10

0.97    0.98    0.97    update-index_range=100

0.98    0.99    1.04    update-inlist_range=100

0.98    0.99    0.98    update-nonindex_range=100

0.99    0.99    0.98    update-one_range=100

0.98    0.99    0.98    update-zipf_range=100

0.99    0.97    0.95    write-only_range=10000

Results: Postgres 12 to 18

All files are here.

Results per microbenchmark from vmstat and iostat are here for the first and second run.

The data below with a larger font is here.

Some comments:

• the hot-points test is almost 2X faster starting in 17.6
• scan is ~1.2X faster starting in 14.19
• all write tests are much faster staring in 17.6

Relative to: 12.22

col-1 : 13.22

col-2 : 14.19

col-3 : 15.14

col-4 : 16.10

col-5 : 17.6

col-6 : 18.0 with the x10b config

col-7 : 18.0 with the x10c config

col-8 : 18.0 with the x10d config

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   point queries, first run

1.02    1.00    1.01    1.00    1.94    1.87    1.91    1.82    hot-points_range=100

1.01    1.02    1.02    1.00    1.02    0.99    1.00    0.98    point-query_range=100

1.02    1.02    1.01    1.03    1.01    1.01    1.00    1.00    points-covered-pk_range=100

1.01    1.04    1.03    1.05    1.03    1.02    1.03    1.03    points-covered-si_range=100

1.01    1.01    1.01    1.02    1.02    1.00    1.00    1.00    points-notcovered-pk_range=100

1.00    1.03    1.02    1.03    1.02    1.01    1.01    1.02    points-notcovered-si_range=100

1.01    1.02    1.02    1.03    1.00    1.00    1.00    0.99    random-points_range=1000

1.01    1.02    1.02    1.02    1.02    1.00    1.00    1.00    random-points_range=100

1.02    1.03    1.02    1.02    1.01    1.00    1.00    1.00    random-points_range=10

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   point queries, second run

1.00    0.98    0.99    1.00    1.94    1.90    1.93    1.92    hot-points_range=100

1.00    1.01    1.02    1.03    1.03    1.02    1.02    1.02    point-query_range=100

1.02    1.01    1.00    1.04    0.99    1.00    1.00    0.99    points-covered-pk_range=100

1.01    1.04    1.03    1.07    1.03    1.03    1.05    1.04    points-covered-si_range=100

1.01    1.02    1.03    1.04    1.01    1.00    0.99    1.01    points-notcovered-pk_range=100

1.02    1.05    1.05    1.05    1.03    1.03    1.03    1.04    points-notcovered-si_range=100

1.01    1.02    1.03    1.03    0.99    0.99    1.00    1.00    random-points_range=1000

1.02    1.02    1.03    1.04    1.01    1.01    1.00    1.01    random-points_range=100

1.02    1.02    1.02    1.03    1.02    1.01    1.01    1.02    random-points_range=10

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   range queries without aggregation, first run

1.00    1.02    1.02    1.01    1.00    0.97    0.98    0.95    range-covered-pk_range=100

1.00    1.02    1.02    1.01    1.00    0.97    0.97    0.94    range-covered-si_range=100

1.01    1.00    1.00    1.00    0.99    0.97    0.97    0.97    range-notcovered-pk_range=100

0.99    1.00    1.00    0.99    1.01    1.00    1.00    0.99    range-notcovered-si_range=100

0.98    1.24    1.11    1.13    1.16    1.12    1.14    1.11    scan_range=100

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   range queries without aggregation, second run

1.01    1.02    1.02    1.02    1.01    1.00    1.00    0.99    range-covered-pk_range=100

1.01    1.03    1.02    1.02    1.01    1.00    1.01    1.00    range-covered-si_range=100

1.00    0.99    1.00    1.00    0.99    0.97    0.98    0.98    range-notcovered-pk_range=100

1.00    1.00    1.00    0.98    1.01    1.00    1.01    1.01    range-notcovered-si_range=100

1.00    1.27    1.15    1.15    0.97    1.20    1.20    1.18    scan_range=100

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   range queries with aggregation, first run

1.02    1.00    1.00    1.01    0.97    0.96    0.97    0.97    read-only-count_range=1000

1.00    1.00    1.02    1.02    0.98    0.99    0.99    0.99    read-only-distinct_range=1000

1.01    1.00    1.03    1.03    1.00    1.01    1.01    1.01    read-only-order_range=1000

1.00    0.98    1.00    1.06    0.95    0.99    0.99    0.99    read-only_range=10000

1.00    1.00    1.00    1.00    1.00    0.98    0.98    0.98    read-only_range=100

1.00    1.01    1.01    1.00    1.01    0.98    0.99    0.98    read-only_range=10

1.01    1.00    1.02    1.01    1.00    0.99    0.98    0.98    read-only-simple_range=1000

1.00    1.00    1.01    1.00    0.99    0.98    0.98    0.98    read-only-sum_range=1000

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   range queries with aggregation, second run

1.03    1.02    1.02    1.03    0.97    0.97    0.97    0.98    read-only-count_range=1000

1.00    0.99    1.02    1.02    0.98    0.99    0.99    0.99    read-only-distinct_range=1000

1.00    0.99    1.02    1.04    1.02    1.01    1.01    1.02    read-only-order_range=1000

1.01    1.03    1.03    1.06    0.97    0.99    0.99    0.99    read-only_range=10000

0.99    1.00    1.00    1.01    1.00    0.99    0.99    0.99    read-only_range=100

0.99    1.00    1.00    1.00    1.01    0.99    1.00    0.99    read-only_range=10

1.00    0.99    1.01    1.00    0.99    0.98    0.98    0.99    read-only-simple_range=1000

1.00    1.00    1.01    1.01    0.99    0.98    0.98    0.98    read-only-sum_range=1000

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   writes, first run

1.00    1.08    1.08    1.05    1.25    1.24    1.23    1.20    delete_range=100

1.01    1.05    1.04    1.03    1.07    1.06    1.02    1.05    insert_range=100

1.00    1.06    1.07    1.07    1.10    1.09    1.08    1.07    read-write_range=100

1.00    1.07    1.08    1.07    1.13    1.13    1.11    1.11    read-write_range=10

0.99    1.04    1.04    0.90    1.43    1.43    1.41    1.43    update-index_range=100

1.00    1.09    1.08    1.08    1.11    1.15    1.06    1.12    update-inlist_range=100

1.00    1.05    1.05    1.04    1.35    1.34    1.34    1.35    update-nonindex_range=100

1.02    0.95    0.96    0.93    1.19    1.19    1.19    1.20    update-one_range=100

1.00    1.05    1.08    1.07    1.23    1.21    1.22    1.23    update-zipf_range=100

1.01    1.06    1.05    1.01    1.25    1.22    1.20    1.24    write-only_range=10000

col-1   col-2   col-3   col-4   col-5   col-6   col-7   col-8   writes, second run

1.00    1.06    1.07    1.07    1.26    1.23    1.23    1.24    delete_range=100

1.03    1.07    1.05    1.05    1.09    1.07    1.08    1.09    insert_range=100

1.01    1.07    1.08    1.07    1.11    1.10    1.10    1.09    read-write_range=100

0.99    1.04    1.06    1.07    1.13    1.11    1.11    1.12    read-write_range=10

0.99    1.02    1.04    0.87    1.44    1.40    1.41    1.40    update-index_range=100

1.00    1.11    1.12    1.09    1.17    1.14    1.16    1.22    update-inlist_range=100

1.01    1.04    1.06    1.03    1.36    1.33    1.35    1.34    update-nonindex_range=100

1.01    0.95    0.98    0.94    1.22    1.21    1.21    1.20    update-one_range=100

0.99    1.05    1.07    1.07    1.24    1.21    1.22    1.21    update-zipf_range=100

1.02    1.06    1.06    1.02    1.27    1.25    1.23    1.21    write-only_range=10000

My time at Oracle: functional and design specification reviews

I worked at Oracle from 1997 to 2005 for 3 years on the app server team in Portland and the last 5 on DBMS query execution in Redwood Shores. I had a good time there, made many friends and learned a lot.

They had an excellent process for functional and design specification reviews. Like many, I am wary of (too much) process but this wasn't too much. It was just enough.

At a high level, you would write and then get a review for the functional spec. The review was an in-person meeting. Once that was resolved the process would repeat for the design spec. You were expected to write a good spec -- it was better for one person (the author) to spend much time on it to avoid wasting time for the many readers. Many specs would be revisited long after the review because there is turnover and specs are easier to read than source code.

We used FrameMaker to write the specs on Solaris workstations. That was a long time ago. The functional spec I wrote for IEEE754 datatypes was more than 50 pages because I had to document every aspect of PL/SQL and SQL that would be impacted by it (there were so many functions to document). The design spec I wrote for a new sort algorithm was also quite long because I had already implemented the algorithm to collect performance results to justify the effort. The patent attorney copied much of that design doc into the patent resulting in a patent that might be more readable than average.

For each specification you setup a meeting a few weeks out and shared the spec with people who might attend the meeting. In many cases there was feedback via email or in person prior to the meeting that could be resolved before the meeting. But in some cases there was feedback that wouldn't get resolved until the meeting.

It is important to split the functional and design specs, and their reviews. It helps with efficiency and the design review might change a lot based on the outcome of the functional spec review.

There are a variety of responses to the feedback, and all of that was added to an appendix of the spec (both the feedback and the response). Common responses include:

• good point
• I will change my spec as you suggest
• no thank you
• I disagree and will not change my spec as you suggest. Hopefully this isn't the response to all feedback but some people like to bike shed and/or get in the way of progress. When I rewrote the sort algorithm, I used something that was derived from quicksort and quicksort implementations have worse than expected performance on some input sequences. The algorithm I used was far better than vanilla quicksort in that regard, but it didn't eliminate the risk. However, the performance improvement over the existing code was so large (the white paper claims 5X faster) that I sad no thank you and the project got done. But I did spend some time doing the math to show how likely (or unlikely) the worst cases were. I needed a tool with arbitrary precision math to for that because the numbers are small and might have ended up using a Scheme implementation.
• good point, but
• I won't change my spec, but I have a workaround for the problem you mention. For IEEE754 datatypes, a few people objected because a few infrequently and fading platforms for the DBMS did not have hardware support for IEEE754. My solution was to use functions for each IEEE754 operation that were trivial for platforms with IEEE754 HW support -- things like double multiply_double(x, y) { return x*y } but could be implemented as needed on the platforms that lacked IEEE754 via a software implementation of IEEE754.

Measuring scaleup for Postgres 18.0 with sysbench

This post has results to measure scaleup for Postgres 18.0 on a 48-core server.

tl;dr

• Postgres continues to be boring (in a good way)
• Results are mostly excellent
• A few of the range query tests have a scaleup that is less than great but I need time to debug

Builds, Configuration & Hardware

The server has an AMD EPYC 9454P 48-Core Processor with AMD SMT disabled, 128G of RAM and SW RAID 0 with 2 NVMe devices. The OS is Ubuntu 22.04.

I compiled Postgres 18.0 from source and the configuration file is 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 Postgres. Each microbenchmark is run for 300 seconds.

The benchmark is run with 1, 2, 4, 8, 12, 16, 20, 24, 32, 40 and 48 clients. The purpose is to determine how well Postgres scales up.

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 still use relative QPS here, but in a different way. The relative QPS here is:

(QPS at X clients) / (QPS at 1 client)

The goal is to determine scaleup efficiency for Postgres. When the relative QPS at X clients is a value near X, then things are great. But sometimes things aren't great and the relative QPS is much less than X. One issue is data contention for some of the write-heavy microbenchmarks. Another issue is mutex and rw-lock contention.

Perf debugging via vmstat and iostat

I use normalized results from vmstat and iostat to help explain why things aren't as fast as expected. By normalized I mean I divide the average values from vmstat and iostat by QPS to see things like how much CPU is used per query or how many context switches occur per write. And note that a high context switch rate is often a sign of mutex contention.

Those results are here but can be difficult to read.

Charts: point queries

The spreadsheet with all of the results is here.

While results aren't perfect, they are excellent. Perfect results would be to get a scaleup of 48 at 48 clients and here the result is between 40 and 42 in most tests. The worst-case is for hot-points where the scaleup is 32.57 at 48 clients. Note that the hot-points test has the most data contention of the point-query tests, as all queries fetch the same rows.

From the vmstat metrics (see here) I don't see an increase in mutex contention (more context switches, see the cs/o column) but I do see an increase in CPU (cpu/o). When compared to a test that has better scaleup, like points-covered-pk, there I also don't see an increase in mutex contention and do see an increase in CPU overhead (see cpu/o) but the CPU increase is smaller (see here).

Charts: range queries without aggregation

The spreadsheet with all of the results is here.

The results again are great, but not perfect. The worst case is for range-notcovered-pk where the scaleup is 32.92 at 48 clients. The base case is for scan where the scaleup is 46.56 at 48 clients.

From the vmstat metrics for range-notcovered-pk I don't see any obvious problems. The CPU overhead (cpu/o, CPU per query) increases by 1.08 (about 8%) from 1 to 48 clients while the context switches per query (cs/o) decreases (see here).

Charts: range queries with aggregation

The spreadsheet with all of the results is here.

Results for range queries with aggregation are worse than for range queries without aggregation. I hope to try and explain that later. A perfect result is scaleup equal to 48. Here, 3 of 8 tests have scaleup less than 3, 4 have scaleup between 30 and 40, and the best case is read-only_range=10 with a scaleup of 43.35.

The worst-case was read-only-count with a scaleup of 21.38. From the vmstat metrics I see that at CPU overhead (cpu/o, CPU per query) increases by 2.08 at 48 clients vs 1 client while context switches per query (cs/o) decrease (see here). I am curious about that CPU increase as isn't as bad for the other range query tests, for example see here where it is no larger than 1.54. The query for read-only-count is here.

Later I hope to explain why read-only-count, read-only-simple and read-only-sum don't do better.

Charts: writes

The spreadsheet with all of the results is here.

The worst-case is update-one where scaleup is 2.86 at 48 clients. The bad result is expected as having many concurrent clients update the same row is an anti-pattern with Postgres. The scaleup for Postgres on that test is a lot worse than for MySQL where it was ~8 with InnoDB. But I am not here for Postgres vs InnoDB arguments.

Excluding the tests that mix reads and writes (read-write-*) the scaleup is between 13 and 21. This is far from great but isn't horrible. I run with fsync-on-commit disabled which highlights problems but is less realistic. So for now I am happy with this results.

Measuring scaleup for MariaDB with sysbench

This post has results to measure scaleup for MariaDB 11.8.3 on a 48-core server.

tl;dr

• Scaleup is better for range queries than for point queries
• For tests where results were less than great, the problem appears to be mutex contention within InnoDB

Builds, Configuration & Hardware

The server has an AMD EPYC 9454P 48-Core Processor with AMD SMT disabled, 128G of RAM and SW RAID 0 with 2 NVMe devices. The OS is Ubuntu 22.04.

I compiled MariaDB 11.8.3 from source and the my.cnf file is 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 MariaDB. Each microbenchmark is run for 300 seconds.

The benchmark is run with 1, 2, 4, 8, 12, 16, 20, 24, 32, 40 and 48 clients. The purpose is to determine how well MariaDB scales up.

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 still use relative QPS here, but in a different way. The relative QPS here is:

(QPS at X clients) / (QPS at 1 client)

The goal is to determine scaleup efficiency for MariaDB. When the relative QPS at X clients is a value near X, then things are great. But sometimes things aren't great and the relative QPS is much less than X. One issue is data contention for some of the write-heavy microbenchmarks. Another issue is mutex and rw-lock contention.

Perf debugging via vmstat and iostat

I use normalized results from vmstat and iostat to help explain why things aren't as fast as expected. By normalized I mean I divide the average values from vmstat and iostat by QPS to see things like how much CPU is used per query or how many context switches occur per write. And note that a high context switch rate is often a sign of mutex contention.

Charts: point queries

The spreadsheet with all of the results is here.

For point queries

• tests for which the relative QPS at 48 clients is greater than 40
• point-query
• tests for which the relative QPS at 48 clients is between 30 and 40
• none
• tests for which the relative QPS at 48 clients is between 20 and 30
• hot-points, points-covered-si, random-points_range=10
• tests for which the relative QPS at 48 clients is between 10 and 20
• points-covered-pk, points-notcovered-pk, points-notcovered-si, random-points_range=100
• tests for which the relative QPS at 48 clients is less than 10
• random-points_range=1000

For 5 of the 9 point query tests, QPS stops improving beyond 16 clients. And I assume that mutex contention is the problem.

Results for the random-points_range=Z tests are interesting. They use oltp_inlist_select.lua which does a SELECT with a large IN-list where the IN-list entries can find rows by exact match on the PK. The value of Z is the number of entries in the IN-list. And here MariaDB scales worse with a larger Z (1000) than with a smaller Z (10 or 100), which means that the thing that limits scaleup is more likely in InnoDB than the parser or optimizer.

From the normalized vmstat metrics (see here) for 1 client and 48 clients the number of context switches per query (the cs/o column) grows a lot more from 1 to 48 clients for random-points_range=1000 than for random-points_range=10. The ratio (cs/o at 48 clients / cs/o at 1 client) is 1.46 for random-points_range=10 and then increases to 19.96 for random-points_range=1000. The problem appears to be mutex contention.

Charts: range queries without aggregation

The spreadsheet with all of the results is here.

For range queries without aggregation:

• tests for which the relative QPS at 48 clients is greater than 40
• range-covered-pk, range-covered-si, range-notcovered-pk
• tests for which the relative QPS at 48 clients is between 30 and 40
• scan
• tests for which the relative QPS at 48 clients is between 20 and 30
• none
• tests for which the relative QPS at 48 clients is between 10 and 20
• none
• tests for which the relative QPS at 48 clients is less than 10
• range-notcovered-si

Only one test has less than great results for scaleup -- range-notcovered-si. QPS for it stops growing beyond 12 clients. The root cause appears to be mutex contention based on the large value for cs/o in the normalized vmstat metrics (see here). For all of the range-*covered-* tests, has the most InnoDB activity per query -- the query isn't covering so it must do PK index access per index entry it finds in the secondary index.

Charts: range queries with aggregation

The spreadsheet with all of the results is here.

For range queries with aggregation:

• tests for which the relative QPS at 48 clients is greater than 40
• read-only-distinct, read-only-order, read-only-range=Y, read-only-sum
• tests for which the relative QPS at 48 clients is between 30 and 40
• read-only-count, read-only-simple
• tests for which the relative QPS at 48 clients is between 20 and 30
• none
• tests for which the relative QPS at 48 clients is between 10 and 20
• none
• tests for which the relative QPS at 48 clients is less than 10
• none

Results here are excellent, and better than the results above for range queries without aggregation. The difference might mean that there is less concurrent activity within InnoDB because aggregation code is run after each row is fetched from InnoDB.

Charts: writes

The spreadsheet with all of the results is here.

For writes:

• tests for which the relative QPS at 48 clients is greater than 40
• none
• tests for which the relative QPS at 48 clients is between 30 and 40
• read-write_range=Y
• tests for which the relative QPS at 48 clients is between 20 and 30
• update-index, write-only
• tests for which the relative QPS at 48 clients is between 10 and 20
• delete, insert, update-inlist, update-nonindex, update-zipf
• tests for which the relative QPS at 48 clients is less than 10
• update-one

The best result is for the read-write_range=Y tests which are the classic sysbench transaction that does a mix of writes, point and range queries. 

The worst result is from update-one which suffers from data contention as all updates are to the same row. A poor result is expected here.

Postgres 18.0 vs sysbench on a 24-core, 2-socket server

This post has results from sysbench run at higher concurrency for Postgres versions 12 through 18 on a server with 24 cores and 2 sockets. My previous post had results for sysbench run with low concurrency. The goal is to search for regressions from new CPU overhead and mutex contention.

tl;dr, from Postgres 17.6 to 18.0

• For most microbenchmarks Postgres 18.0 is between 1% and 3% slower than 17.6
• The root cause might be new CPU overhead. It will take more time to gain confidence in results like this. On other servers with sysbench run at low concurrency I only see regressions for some of the range-query microbenchmarks. Here I see them for point-query and writes.

tl;dr, from Postgres 12.22 through 18.0

• For point queries Postgres 18.0 is usually about 5% faster than 12.22
• For range queries Postgres 18.0 is usually as fast as 12.22
• For writes Postgres 18.0 is much faster than 12.22

Builds, configuration and hardware

I compiled Postgres from source for versions 12.22, 13.22, 14.19, 15.14, 16.10, 17.6, and 18.0.

The server is a SuperMicro SuperWorkstation 7049A-T with 2 sockets, 12 cores/socket, 64G RAM. The CPU is Intel Xeon Silver 4214R CPU @ 2.40GHz. It runs Ubuntu 24.04. Storage is a 1TB m.2 NVMe device with ext-4 and discard enabled.

Prior to 18.0, the configuration file was named conf.diff.cx10a_c24r64 and is here for 12.22, 13.22, 14.19, 15.14, 16.10 and 17.6.

For 18.0 I tried 3 configuration files:

• conf.diff.cx10b_c24r64 (x10b) - uses io_method=sync
• conf.diff.cx10c_c24r64 (x10c) - uses io_method=worker
• conf.diff.cx10d_c24r64 (x10d) - uses 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 read-heavy microbenchmarks run for 600 seconds and the write-heavy for 900 seconds.

The benchmark is run with 16 clients and 8 tables with 10M rows per table. The purpose is to search for regressions from new CPU overhead and mutex contention.

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 base version)

When the relative QPS is > 1 then some version is faster than base version.  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.

I present results for:

• versions 12 through 18 using 12.22 as the base version
• versions 17.6 and 18.0 using 17.6 as the base version

Results: Postgres 17.6 and 18.0

Results per microbenchmark from vmstat and iostat are here.

For point queries, 18.0 often gets between 1% and 3% less QPS than 17.6 and the root cause might be new CPU overhead. See the cpu/o column (CPU per query) in the vmstat metrics here for the random-points microbenchmarks.

For range queries, 18.0 often gets between 1% and 3% less QPS than 17.6 and the root cause might be new CPU overhead. See the cpu/o column (CPU per query) in the vmstat metrics here for the read-only_range=X microbenchmarks.

For writes queries, 18.0 often gets between 1% and 2% less QPS than 17.6 and the root cause might be new CPU overhead. I ignore the write-heavy microbenchmarks that also do queries as the regressions for them might be from the queries. See the cpu/o column (CPU per query) in the vmstat metrics here for the update-index microbenchmark.

Relative to: 17.6

col-1 : 18.0 with the x10b config

col-2 : 18.0 with the x10c config

col-3 : 18.0 with the x10d config

col-1   col-2   col-3   point queries

1.00    0.99    1.00    hot-points_range=100

0.99    0.98    1.00    point-query_range=100

0.98    0.99    0.99    points-covered-pk_range=100

0.99    0.99    0.98    points-covered-si_range=100

0.97    0.99    0.98    points-notcovered-pk_range=100

0.98    0.99    0.97    points-notcovered-si_range=100

0.98    0.99    0.98    random-points_range=1000

0.97    0.99    0.98    random-points_range=100

0.99    0.99    0.98    random-points_range=10

col-1   col-2   col-3   range queries without aggregation

0.98    0.98    0.99    range-covered-pk_range=100

0.98    0.98    0.98    range-covered-si_range=100

0.98    0.99    0.98    range-notcovered-pk_range=100

1.00    1.02    0.99    range-notcovered-si_range=100

1.01    1.01    1.01    scan_range=100

col-1   col-2   col-3   range queries with aggregation

0.99    1.00    0.98    read-only-count_range=1000

0.98    0.98    0.98    read-only-distinct_range=1000

0.97    0.97    0.96    read-only-order_range=1000

0.97    0.98    0.97    read-only_range=10000

0.98    0.99    0.98    read-only_range=100

0.99    0.99    0.99    read-only_range=10

0.98    0.99    0.99    read-only-simple_range=1000

0.98    1.00    0.98    read-only-sum_range=1000

col-1   col-2   col-3   writes

0.99    0.99    0.99    delete_range=100

0.99    0.99    0.99    insert_range=100

0.98    0.98    0.98    read-write_range=100

0.99    1.00    0.99    read-write_range=10

0.99    0.98    0.97    update-index_range=100

0.99    0.99    1.00    update-inlist_range=100

1.00    0.97    0.99    update-nonindex_range=100

0.97    1.00    0.98    update-one_range=100

1.00    0.99    1.01    update-zipf_range=100

0.98    0.98    0.97    write-only_range=10000

Results: Postgres 12 to 18

For the Postgres 18.0 results in col-6, the result is in green when relative QPS is >= 1.05 and in yellow when relative QPS is <= 0.98. Yellow indicates a possible regression.

Results per microbenchmark from vmstat and iostat are here.

Relative to: 12.22

col-1 : 13.22

col-2 : 14.19

col-3 : 15.14

col-4 : 16.10

col-5 : 17.6

col-6 : 18.0 with the x10b config

col-1   col-2   col-3   col-4   col-5   col-6   point queries

0.98    0.96    0.99    0.98    2.13    2.13    hot-points_range=100

1.00    1.02    1.01    1.02    1.03    1.01    point-query_range=100

0.99    1.05    1.05    1.08    1.07    1.05    points-covered-pk_range=100

0.99    1.08    1.05    1.07    1.07    1.05    points-covered-si_range=100

0.99    1.04    1.05    1.06    1.07    1.05    points-notcovered-pk_range=100

0.99    1.05    1.04    1.05    1.06    1.04    points-notcovered-si_range=100

0.98    1.03    1.04    1.06    1.06    1.04    random-points_range=1000

0.98    1.04    1.05    1.07    1.07    1.05    random-points_range=100

0.99    1.02    1.03    1.05    1.05    1.04    random-points_range=10

col-1   col-2   col-3   col-4   col-5   col-6   range queries without aggregation

1.02    1.04    1.03    1.04    1.03    1.01    range-covered-pk_range=100

1.05    1.07    1.06    1.06    1.06    1.05    range-covered-si_range=100

0.99    1.00    1.00    1.00    1.01    0.98    range-notcovered-pk_range=100

0.97    0.99    1.00    1.01    1.01    1.01    range-notcovered-si_range=100

0.86    1.06    1.08    1.17    1.18    1.20    scan_range=100

col-1   col-2   col-3   col-4   col-5   col-6   range queries with aggregation

0.98    0.97    0.97    1.00    0.98    0.97    read-only-count_range=1000

0.99    0.99    1.02    1.02    1.01    0.99    read-only-distinct_range=1000

1.00    0.99    1.02    1.05    1.05    1.02    read-only-order_range=1000

0.99    0.99    1.04    1.07    1.09    1.06    read-only_range=10000

0.99    1.00    1.00    1.01    1.02    0.99    read-only_range=100

1.00    1.00    1.00    1.01    1.01    1.00    read-only_range=10

0.99    0.99    1.00    1.01    1.01    0.99    read-only-simple_range=1000

0.98    0.99    0.99    1.00    1.00    0.98    read-only-sum_range=1000

col-1   col-2   col-3   col-4   col-5   col-6   writes

0.98    1.09    1.09    1.04    1.29    1.27    delete_range=100

0.99    1.03    1.02    1.03    1.08    1.07    insert_range=100

1.00    1.03    1.04    1.05    1.07    1.05    read-write_range=100

1.01    1.09    1.09    1.09    1.15    1.14    read-write_range=10

1.00    1.04    1.03    0.86    1.44    1.42    update-index_range=100

1.01    1.11    1.11    1.12    1.13    1.12    update-inlist_range=100

0.99    1.04    1.06    1.05    1.25    1.25    update-nonindex_range=100

1.05    0.92    0.90    0.84    1.18    1.15    update-one_range=100

0.98    1.04    1.03    1.01    1.26    1.26    update-zipf_range=100

1.02    1.05    1.10    1.09    1.21    1.18    write-only_range=10000