Fast Imaging for Webserver

[Prateek]

Hi. I'm creating a web-application using CherryPy 2.2.1. My application
needs to process images (JPG/PNG files) to

1) create thumbnails (resize them)
2) overlay them on a custom background (a simple frame)
3) Overlay 'badges' (small 16x16 images) on top of the final thumbnail

I am using PIL 1.1.5 which I have custom compiled on my development
machine (MacBook Pro 2.33Ghz Core Duo)..
I am using im.thumbnail for step 1 and im.paste for steps 2 and 3.

The problem is that this thing is just way too slow.
For ab -n 1000 -C session_id=2f55ae2dfefa896f67a80f73045aadfa4b4269f 1
http://localhost:8080/imaging/icon/def/128/255 (where def is the name
of the image - default in this case - 128 is the size in pixels and 255
is the background color), I am getting:

Document Path: /imaging/icon/def/128/255
Document Length: 14417 bytes

Concurrency Level: 1
Time taken for tests: 18.664 seconds
Complete requests: 1000
Failed requests: 0
Broken pipe errors: 0
Total transferred: 14680000 bytes
HTML transferred: 14417000 bytes
Requests per second: 53.58 [#/sec] (mean)
Time per request: 18.66 [ms] (mean)
Time per request: 18.66 [ms] (mean, across all concurrent
requests)
Transfer rate: 786.54 [Kbytes/sec] received

FYI: This request returns a PNG image (image/png) and not html

My understanding is that the problem is either with the CherryPy setup
(which is likely because even in other cases, i don't get much more
than 65 requests per second) or PIL itself (even though I'm caching the
background images and source images)

Does anyone have a better solution? Is there a faster replacement for
PIL?

Thanks in advance.
Prateek

[Paul McGuire]

On Jan 25, 12:46 am, "Prateek" <[e-mail]:

Concurrency Level: 1
Time taken for tests: 18.664 seconds
Complete requests: 1000
Failed requests: 0
Broken pipe errors: 0
Total transferred: 14680000 bytes
HTML transferred: 14417000 bytes
Requests per second: 53.58 [#/sec] (mean)
Time per request: 18.66 [ms] (mean)
Time per request: 18.66 [ms] (mean, across all concurrent
requests)
Transfer rate: 786.54 [Kbytes/sec] received

FYI: This request returns a PNG image (image/png) and not html

My understanding is that the problem is either with the CherryPy setup
(which is likely because even in other cases, i don't get much more
than 65 requests per second) or PIL itself (even though I'm caching the
background images and source images)

Does anyone have a better solution? Is there a faster replacement for
PIL?

So you have some gross level statistics on how long a request takes to
process, that's good. But before you start replacing PIL, or
optimizing CherryPy, or other possible performance-improving efforts,
you should profile the within-request processing, find the bottleneck,
and take care of that first. Without profiling info, you can easily
spend days optimizing out 90% of a task that takes 2% of the total
processing time, for little net gain.

-- Paul

[half.italian]

On Jan 24, 10:46 pm, "Prateek" <[e-mail]:

Hi. I'm creating a web-application using CherryPy 2.2.1. My application
needs to process images (JPG/PNG files) to

1) create thumbnails (resize them)
2) overlay them on a custom background (a simple frame)
3) Overlay 'badges' (small 16x16 images) on top of the final thumbnail

I am using PIL 1.1.5 which I have custom compiled on my development
machine (MacBook Pro 2.33Ghz Core Duo)..
I am using im.thumbnail for step 1 and im.paste for steps 2 and 3.

The problem is that this thing is just way too slow.
For ab -n 1000 -C session_id=2f55ae2dfefa896f67a80f73045aadfa4b4269f 1http://localhost:8080/imaging/icon/def/128/255(where def is the name
of the image - default in this case - 128 is the size in pixels and 255
is the background color), I am getting:

Document Path: /imaging/icon/def/128/255
Document Length: 14417 bytes

Concurrency Level: 1
Time taken for tests: 18.664 seconds
Complete requests: 1000
Failed requests: 0
Broken pipe errors: 0
Total transferred: 14680000 bytes
HTML transferred: 14417000 bytes
Requests per second: 53.58 [#/sec] (mean)
Time per request: 18.66 [ms] (mean)
Time per request: 18.66 [ms] (mean, across all concurrent
requests)
Transfer rate: 786.54 [Kbytes/sec] received

FYI: This request returns a PNG image (image/png) and not html

My understanding is that the problem is either with the CherryPy setup
(which is likely because even in other cases, i don't get much more
than 65 requests per second) or PIL itself (even though I'm caching the
background images and source images)

Does anyone have a better solution? Is there a faster replacement for
PIL?

Thanks in advance.
Prateek

Check out the mac CoreGraphics bindings. Not sure if it will be any
faster, but it can do what you want.

http://www.macdevcenter.com/pub/a/ma..._graphics.html

~Sean DiZazzo

[bearophileHUGS]

Paul McGuire:

before you start replacing PIL, or
optimizing CherryPy, or other possible performance-improving efforts,
you should profile the within-request processing, find the bottleneck,
and take care of that first.

Good advice.
Among the tests, the OP can also try to change the antialiasing type of
PIL, going to the nearest neighbor, and doing some transfer timings
again. It may help find where the bottleneck is.

Bye,
bearophile

[Fredrik Lundh]

"Prateek" wrote:

Hi. I'm creating a web-application using CherryPy 2.2.1. My application
needs to process images (JPG/PNG files) to

1) create thumbnails (resize them)
2) overlay them on a custom background (a simple frame)
3) Overlay 'badges' (small 16x16 images) on top of the final thumbnail

I am using PIL 1.1.5 which I have custom compiled on my development
machine (MacBook Pro 2.33Ghz Core Duo)..
I am using im.thumbnail for step 1 and im.paste for steps 2 and 3.

The problem is that this thing is just way too slow.
For ab -n 1000 -C session_id=2f55ae2dfefa896f67a80f73045aadfa4b4269f 1
http://localhost:8080/imaging/icon/def/128/255 (where def is the name
of the image - default in this case - 128 is the size in pixels and 255
is the background color), I am getting:

Document Path: /imaging/icon/def/128/255
Document Length: 14417 bytes

Concurrency Level: 1
Time taken for tests: 18.664 seconds
Complete requests: 1000
Failed requests: 0
Broken pipe errors: 0
Total transferred: 14680000 bytes
HTML transferred: 14417000 bytes
Requests per second: 53.58 [#/sec] (mean)
Time per request: 18.66 [ms] (mean)
Time per request: 18.66 [ms] (mean, across all concurrent
requests)
Transfer rate: 786.54 [Kbytes/sec] received

FYI: This request returns a PNG image (image/png) and not html

My understanding is that the problem is either with the CherryPy setup
(which is likely because even in other cases, i don't get much more
than 65 requests per second) or PIL itself (even though I'm caching the
background images and source images)

without knowing more about what kind of source images you're using, the
amount of data they represent etc, it's not entirely obvious to me that 50+
images per second on a single server is that bad, really. unless the source
images are rather tiny, that's plenty of pixels per second to crunch.

are you really receiving more than 4500000 source images per day per
server?

</F>

[Prateek]

Thanks for all the responses so far..

The 'imaging' app that I mentioned is actually serving as a library for
a bunch of other apps.
A typical scenario is icon processing. The source image *could* be
large (in this case it was 128x128 PNGA)

Think of it as a Mac Finder/Win Explorer style view (grid + imagelist)
where you're showing a bunch of data in icon form (16x16). For small
amounts of data, everything is ok... but as soon as I get to 600-700
items, the page takes forever to load. The obvious solution is
pagination but I'm doing some aggregate analysis on the data (like a
frequency distribution) where if I were to paginate the data, i'd lose
the client-side sorting ability (I'd have to sort server side which is
a whole other can of worms).

I'm gonna take Paul's advice and report as soon as I can. One other
alternative is to save this information on the filesystem (as temp
files) and route all future requests via the static file processing
mechanism (which can be handled by Apache)... do you think that is a
good idea?

Prateek

On Jan 25, 3:11 pm, "Fredrik Lundh" <[e-mail]:

"Prateek" wrote:

Hi. I'm creating a web-application using CherryPy 2.2.1. My application
needs to process images (JPG/PNG files) to

1) create thumbnails (resize them)
2) overlay them on a custom background (a simple frame)
3) Overlay 'badges' (small 16x16 images) on top of the final thumbnail

I am using PIL 1.1.5 which I have custom compiled on my development
machine (MacBook Pro 2.33Ghz Core Duo)..
I am using im.thumbnail for step 1 and im.paste for steps 2 and 3.

The problem is that this thing is just way too slow.
For ab -n 1000 -C session_id=2f55ae2dfefa896f67a80f73045aadfa4b4269f 1
http://localhost:8080/imaging/icon/def/128/255(where def is the name
of the image - default in this case - 128 is the size in pixels and 255
is the background color), I am getting:

Document Path: /imaging/icon/def/128/255
Document Length: 14417 bytes

Concurrency Level: 1
Time taken for tests: 18.664 seconds
Complete requests: 1000
Failed requests: 0
Broken pipe errors: 0
Total transferred: 14680000 bytes
HTML transferred: 14417000 bytes
Requests per second: 53.58 [#/sec] (mean)
Time per request: 18.66 [ms] (mean)
Time per request: 18.66 [ms] (mean, across all concurrent
requests)
Transfer rate: 786.54 [Kbytes/sec] received

FYI: This request returns a PNG image (image/png) and not html

My understanding is that the problem is either with the CherryPy setup
(which is likely because even in other cases, i don't get much more
than 65 requests per second) or PIL itself (even though I'm caching the
background images and source images)without knowing more about what kind of source images you're using, the

amount of data they represent etc, it's not entirely obvious to me that 50+
images per second on a single server is that bad, really. unless the source
images are rather tiny, that's plenty of pixels per second to crunch.

are you really receiving more than 4500000 source images per day per
server?

</F>