thumbnail_image is 6x faster than thumbnail for resizing

[LuckyFrost]

In many places I see people saying not to use thumbnail_image because it is very slow. It is even written in the documentation. However, I seem to be making a mistake somewhere, because in my case thumbnail_image is dramatically faster than the other options.

"Test stand":
Requests go through nginx+php-fpm
Oracle Linux Server 8.7
PHP 8.5.5
vips 8.18.2
Xeon(R) CPU E5-2687W v4
Load generator: hey
Test image:
921-020.jpg JPEG 800x800 800x800+0+0 8-bit sRGB 74103B 0.000u 0:00.000

For testing, I wrote a simple script:

<?php
require __DIR__ . '/vendor/autoload.php';
use Jcupitt\Vips\Image;

$icon_h = 330;
$icon_w = 330;
$image_path = '/srv/code/921-020.jpg';
$targetOptions = ['optimize-coding' => true, 'strip' => true, 'Q' => 86];
$thumbnailOptions = ['height' => $icon_h, 'crop' => 'centre', 'size' => 'down'];
$targetSuffix = '.jpg';

header('Content-Type: image/jpeg');

switch ($_GET['bench']) {
  case "thumbnail_image":
    $image = Image::newFromFile($image_path);
    $image = $image->thumbnail_image($icon_w, $thumbnailOptions);
    echo $image->writeToBuffer($targetSuffix,$targetOptions);
  break;
  case "thumbnail":
    $image=Image::thumbnail($image_path, $icon_w, $thumbnailOptions);
    echo $image->writeToBuffer($targetSuffix,$targetOptions);
  break;
  case "thumbnail_buffer":
    $image = file_get_contents($image_path);
    $image=Image::thumbnail_buffer($image, $icon_w, $thumbnailOptions);
    echo $image->writeToBuffer($targetSuffix,$targetOptions);
  break;
  }

Test results:

./hey_linux_amd64 -c 20 -n 20000 "https://mydomain/image.jpg?bench=thumbnail_image"

Summary:
  Total:        7.8730 secs
  Slowest:      0.0264 secs
  Fastest:      0.0039 secs
  Average:      0.0078 secs
  Requests/sec: 2540.3259


Response time histogram:
  0.004 [1]     |
  0.006 [5174]  |■■■■■■■■■■■■■■■■■■■■■■■■
  0.008 [8674]  |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.011 [4024]  |■■■■■■■■■■■■■■■■■■■
  0.013 [1525]  |■■■■■■■
  0.015 [431]   |■■
  0.017 [121]   |■
  0.020 [38]    |
  0.022 [7]     |
  0.024 [3]     |
  0.026 [2]     |


Latency distribution:
  10%% in 0.0057 secs
  25%% in 0.0061 secs
  50%% in 0.0071 secs
  75%% in 0.0089 secs
  90%% in 0.0108 secs
  95%% in 0.0120 secs
  99%% in 0.0149 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0000 secs, 0.0000 secs, 0.0079 secs
  DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0022 secs
  req write:    0.0000 secs, 0.0000 secs, 0.0016 secs
  resp wait:    0.0077 secs, 0.0039 secs, 0.0263 secs
  resp read:    0.0000 secs, 0.0000 secs, 0.0016 secs

Status code distribution:
  [200] 20000 responses

./hey_linux_amd64 -c 20 -n 20000 "https://mydomain/image.jpg?bench=thumbnail"

Summary:
  Total:        49.4876 secs
  Slowest:      0.0964 secs
  Fastest:      0.0232 secs
  Average:      0.0492 secs
  Requests/sec: 404.1420


Response time histogram:
  0.023 [1]     |
  0.030 [26]    |
  0.038 [150]   |■
  0.045 [5215]  |■■■■■■■■■■■■■■■■■■■■■■■
  0.052 [9225]  |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.060 [4184]  |■■■■■■■■■■■■■■■■■■
  0.067 [1008]  |■■■■
  0.074 [166]   |■
  0.082 [20]    |
  0.089 [3]     |
  0.096 [2]     |


Latency distribution:
  10%% in 0.0419 secs
  25%% in 0.0448 secs
  50%% in 0.0485 secs
  75%% in 0.0529 secs
  90%% in 0.0575 secs
  95%% in 0.0605 secs
  99%% in 0.0670 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0000 secs, 0.0000 secs, 0.0063 secs
  DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0011 secs
  req write:    0.0000 secs, 0.0000 secs, 0.0016 secs
  resp wait:    0.0491 secs, 0.0231 secs, 0.0963 secs
  resp read:    0.0001 secs, 0.0000 secs, 0.0018 secs

Status code distribution:
  [200] 20000 responses

./hey_linux_amd64 -c 20 -n 20000 "https://mydomain/image.jpg?bench=thumbnail_buffer"

Summary:
  Total:        48.5924 secs
  Slowest:      0.0839 secs
  Fastest:      0.0232 secs
  Average:      0.0484 secs
  Requests/sec: 411.5871


Response time histogram:
  0.023 [1]     |
  0.029 [16]    |
  0.035 [31]    |
  0.041 [1917]  |■■■■■■■■■■
  0.047 [7619]  |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.054 [6772]  |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.060 [2775]  |■■■■■■■■■■■■■■■
  0.066 [711]   |■■■■
  0.072 [137]   |■
  0.078 [15]    |
  0.084 [6]     |


Latency distribution:
  10%% in 0.0415 secs
  25%% in 0.0442 secs
  50%% in 0.0478 secs
  75%% in 0.0519 secs
  90%% in 0.0563 secs
  95%% in 0.0590 secs
  99%% in 0.0651 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0000 secs, 0.0000 secs, 0.0087 secs
  DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0019 secs
  req write:    0.0000 secs, 0.0000 secs, 0.0017 secs
  resp wait:    0.0483 secs, 0.0231 secs, 0.0838 secs
  resp read:    0.0001 secs, 0.0000 secs, 0.0008 secs

Status code distribution:
  [200] 20000 responses

What am I doing wrong?

[jcupitt]

Hello @LuckyFrost,

You're seeing two things:

1. For small images, tricks like shrink-on-load won't have any effect. Try with a larger source image, and try with at least a x2 reduction in size.
2. libvips will cache and reuse results, so your thumbnail_image time is only for encode, not for decode. Try with many different source images, not just one that you reuse.

To see the difference, try making a set of source files:

$ for i in {1..100}; do cp ../k2.jpg $i.jpg; done
$ vipsheader 1.jpg
1.jpg: 1450x2048 uchar, 3 bands, srgb, jpegload

Then try:

$ time for i in {1..100}; do vips thumbnail_image $i.jpg tn_$i.jpg 128; done 

real	0m5.504s
user	0m3.113s
sys	0m2.445s
$ rm tn_*
$ time for i in {1..100}; do vips thumbnail $i.jpg tn_$i.jpg 128; done 

real	0m3.966s
user	0m2.593s
sys	0m1.641s

You can see thumbnail is noticably quicker. You'll see a larger difference with larger source files.

You'll see a large quality improvement with vector formats, like PDF and SVG.

[kleisauke]

It might also be worth testing this change:

@@ -13,7 +13,7 @@ header('Content-Type: image/jpeg');
 
 switch ($_GET['bench']) {
   case "thumbnail_image":
-    $image = Image::newFromFile($image_path);
+    $image = Image::newFromFile($image_path, ['access' => 'sequential']);
     $image = $image->thumbnail_image($icon_w, $thumbnailOptions);
     echo $image->writeToBuffer($targetSuffix,$targetOptions);
   break;

This makes it consistent with the other thumbnail operations, which use sequential access by default. Though, it may slightly impact performance when caching is enabled, as random access mode allows loader caching, whereas sequential mode prevents it.

[LuckyFrost]

Hello.
Thanks for the consultation.
I took 500k images smaller than 1kk pixels and 500k images larger than 1kk pixels and ran a random query bombardment on them, executing at least 100,000 queries for each scenario.

Added env["VIPS_CONCURRENCY"] = 1 to php-fpm conf

Config::cacheSetMax(0);
$icon_h = 330;
$icon_w = 330;
$quality = 90;
$targetOptions = ['optimize-coding' => true, 'strip' => true, 'Q' => $quality, 'profile' => 'srgb'];
$thumbnailOptions = ['height' => $icon_h, 'crop' => 'centre', 'size' => 'down'];

...
  case "q_decrease_vips":
    echo Image::newFromFile($image_path)->jpegsave_buffer($targetOptions);
  break;
  case "thumbnail_image":
    $image = Image::newFromFile($image_path);
    echo $image->thumbnail_image($icon_w, $thumbnailOptions)->jpegsave_buffer($targetOptions);
  break;
  case "thumbnail_image_sequential":
    $image = Image::newFromFile($image_path, ['access' => 'sequential']);
    echo $image->thumbnail_image($icon_w, $thumbnailOptions)->jpegsave_buffer($targetOptions);
  break;
  case "thumbnail":
    echo Image::thumbnail($image_path, $icon_w, $thumbnailOptions)->jpegsave_buffer($targetOptions);
  break;
  case "thumbnail_buffer":
    $image = file_get_contents($image_path);
    echo Image::thumbnail_buffer($image, $icon_w, $thumbnailOptions)->jpegsave_buffer($targetOptions);
  break;

Images greater than 1kk:

PHP
q_decrease_vips
Transaction rate: 207.22 trans/sec

thumbnail_buffer
Transaction rate: 325.18 trans/sec
thumbnail
Transaction rate: 289.52 trans/sec
thumbnail_image_sequential
Transaction rate: 203.21 trans/sec
thumbnail_image
Transaction rate: 192.41 trans/sec

openresty+vips-lua
thumbnail
Transaction rate: 206.19 trans/sec

Lower than 1kk:

PHP
q_decrease_vips
Transaction rate: 686.89 trans/sec

thumbnail_buffer
Transaction rate: 530.50 trans/sec
thumbnail
Transaction rate: 439.74 trans/sec
thumbnail_image_sequential
Transaction rate: 491.11 trans/sec
thumbnail_image
Transaction rate: 494.07 trans/sec

The difference between thumbnail and thumbnail_buffer is a bit surprising, but it looks good either way.