Why is a web server 2-3x slower in WSL than in VMware? (same Docker stack) – PHP Versions

Several caveats with my answer:

• I wish I had a solution for you, but your primary question is why, which this should answer. And perhaps the data here will help lead us to a solution.
• I hope I’m wrong about this, and I may well be. I’ve always heard anecdotally that WSL2 performance was "near native." Your experience, however, coupled with my benchmarking below, leads me to believe that may not be the case.

That said, I’m going to report the data that I came up with in investigating this.

Short summary — My benchmarking seems to show a substantial disk IO and memory performance delta between Hyper-V and VMWare that likely explains your WordPress results.

Supporting data and research

I started out with a similar test scenario as yours, but attempted to reduce it to as much of an MRE as I could:

Configuration

Hardware:

• i9500
• 16GB of RAM
• SSD
• Fairly fresh/recent install of Windows 11 Pro, with WSL2 enabled
• Fresh install of VMWare Workstation 16 Player

Virtualization:

• Default VMWare settings (2 CPUs, 4GB RAM)
• Default WSL2 settings (6 CPUs, 8GB RAM)

In both WSL2 and VMWare:

• Ubuntu Server 20.04 guest/distribution
• Docker installed in both (from official repo, not Docker Desktop)
• ubuntu:latest (22.04) Docker image
• MySQL server (MariaDB) and Sysbench installed in the Ubuntu 22.04 Docker container

Note that for the benchmarking below:

• I closed VMWare when testing WSL2, and vice-versa.
• I did not reboot the Windows host between tests. However, note that some of these tests were run multiple times, from both VMWare and WSL2/Hyper-V with no substantial differences in the results, so I do not believe a reboot would have changed the results measurably.

Benchmarking

I started off with some basic Sysbench testing of CPU and Memory. This was done from within the Docker container.

• A simple sysbench cpu --time=300 run:

  	VMWare	WSL2
  events/sec	1,250.97	1,252.89
  # events	375,294.00	375,869.00
  Latency
  ↪ Min	0.77	0.77
  ↪ Avg	0.80	0.80
  ↪ Max	31.40	4.07
  ↪ 95th percentile	0.87	0.86

  Pretty much an even matchup there.

• sysbench memory run:

  	VMWare	WSL2
  Total operations	64,449,416.00	6,456,274.00
  MiB transferred	62,938.88	6,304.96
  Latency
  ↪ Min	0.00	0.00
  ↪ Avg	0.00	0.00
  ↪ Max	23.63	0.12
  ↪ 95th percentile	0.00	0.00

  Ouch – WSL2’s Docker image is running at about 10% of the memory bandwidth of VMWare. I’ll be honest; it was tough to spot this until I inserted the comma separators here in the table ;-). At first glance, I thought the two were on par.

• I decided to skip straight to MySQL testing, also using Sysbench, since this would probably provide the closest match to your WordPress usage. This was done (after a corresponding prepare) with:

  sysbench oltp_read_write.lua --mysql-user=root --time=300 --tables=10 --table-size=1000000 --range_selects=off --report-interval=1 --histogram run
  

  I’ll skip the histogram and second-by-second results (but I have them saved if they are useful to anyone), but here’s the summary data:

  	VMWare	WSL2
  Queries performed
  ↪ Read	583,220	66,910
  ↪ Write	233,288	26,764
  ↪ Other	116,644	13,382
  ↪ Total	933,152	107,056
  Transactions	58,322	6,691
  Ignored errors	0	0
  Reconnects	0	0
  Latency
  ↪ Min	2.08	14.54
  ↪ Avg	5.14	44.83
  ↪ Max	71.67	193.75
  ↪ 95th Percentile	11.65	81.48

  Again, ouch — WSL2’s MySQL performance (in Docker, at least) is benchmarking at around a tenth of that of VMWare. It’s likely that most of your observed performance difference is represented in these results.

At this point, I began to suspect that the problem could be reproduced in a more generic (IO) way at the hypervisor level, ignoring WSL2 and Docker entirely. WSL2, of course, runs inside a (hidden to the user) Hyper-V backed VM, even though it doesn’t require the full Hyper-V manager.

I proceeded to enable Hyper-V and install another Ubuntu 20.04 guest in it. I then installed Sysbench in both the VMWare and Hyper-V guest Ubuntu OS.

• I then ran a disk IO comparison with:

  sysbench fileio --file-total-size=15G prepare
  sysbench fileio --file-total-size=15G --file-test-mode=rndrw --time=300 --max-requests=0 --histogram run
  

  The results bore out the suspicion:

  	VMWare Ubuntu Guest	Hyper-V Ubuntu Guest
  File operations
  ↪ Reads/sec	2,847.07	258.37
  ↪ Writes/sec	1,898.05	172.25
  ↪ fsyncs/sec	6,074.06	551.20
  Throughput
  ↪ MiB/sec Read	44.49	4.04
  ↪ MiB/sec Written	29.66	2.69
  Latency
  ↪ Min	0.00	0.00
  ↪ Avg	0.09	1.02
  ↪ Max	329.88	82.77
  ↪ 95th Percentile	0.32	4.10

  One interesting thing to note during this stage was that the prepare operation for Sysbench was faster on Hyper-V by about 30% (IIRC). I did not capture the results since the prepare step isn’t supposed to be part of the benchmarking.

  However, after reading your comment and benchmarking results on unzip being faster on WSL2, I’m thinking there may be a connection. Both VMWare and Hyper-V/WSL2 use dynamically resized virtual disks (sometimes called "sparse"). The size of the virtual disk on the host OS essentially starts as a near-0-byte file and grows as needed up to its maximum size.

  It may be that either:

  • Hyper-V has a performance advantage when growing the virtual disk.
  • Or in our testing, VMWare needed to grow the disk for these operations but the Hyper-V/WSL2 disk already had excess free space (from previously deleted files) available.

  I cannot say for sure exactly which order I did things in, and the only way to know for sure would be to "shrink/compress" the virtual disks and try again.

Summary

It appears, to my naïve eye and at least on the "Pro" level of Windows, that Hyper-V has some serious performance limitations when compared to VMWare.

Tuning attempts and other comparisons

I did attempt some tuning of the Hyper-V system, but I’m no expert in that area. Regardless, there’s not much that we as users could do to extend any Hyper-V tuning to WSL2 — Microsoft would have to make most of those changes.

I did try converting the dynamic VHDX to a fixed one, in hopes that it would increase IO, but it did not make a substantial change either.

I’ve also now also tried:

• Disabling swap in WSL2

• Running the Sysbench tests with more threads

• Setting a fixed 12GB RAM size for WSL2.

• Running sysbench under WSL2 on my primary desktop, which has faster memory and an NVMe drive compared to my test system’s SSD.

  • Memory speed was significantly improved – Apples-to-oranges, but my desktop’s memory numbers were comparable to VMWare running on the lower-end test system.
  • However, this made no difference on the disk IO numbers. They were still in the same range as the test system.

Next steps

It would be great if you could, perhaps, run some similar benchmarking apart from your WordPress instance, since you already have the two environments set up. If we can corroborate the data, we should probably report it to, at least, the WSL team. They can hopefully either provide some guidance on how to tune WSL2 to near parity with VMWare, or work with the Hyper-V team on this.

Again, it does seem surprising to me that the disparity between Hyper-V and VMWare is so great. I still struggle to believe that I myself am not doing something wrong in the benchmarking.