Difference between revisions of "Texture Cache"

From Second Life Wiki
Jump to navigation Jump to search
 
(15 intermediate revisions by 7 users not shown)
Line 11: Line 11:
*** Stored in a 'textures' subdirectory in the cache directory, specified in the preferences
*** Stored in a 'textures' subdirectory in the cache directory, specified in the preferences
*** This is a two tiered read/write cache
*** This is a two tiered read/write cache
**** A list of all files in the cace is stored in texture.entries
**** A list of all files in the cache is stored in texture.entries
**** The header for each image (enough to identify the size and load the first mip level) is stored in texture.cache
**** The header for each image (enough to identify the size and load the first mip level) is stored in texture.cache
**** The body (minus the header) for a limited (by the cache size) number of entries is stored in textures/[0-9a-f]/textureid
**** The body (minus the header) for a limited (by the cache size) number of entries is stored in textures/[0-9a-f]/textureid
* Requests to read or write to the cache are processed in the priority given
* Requests to read or write to the cache are processed in the priority given
* The results are given through a [[Responder]] class when they are ready
* The results are given through a [[Responder]] class when they are ready
== Disk storage format ==
* Currently all textures in Second Life are stored in the JPEG2000 compressed format
=== The header consists of two files ===
* cache/texture.entries
''8 byte header:''
<pre>
struct EntriesInfo
{
F32 mVersion;
U32 mEntries;
};
</pre>
:*Version 1.0: ''Array of 24 byte entries''
<pre>
struct Entry
{
LLUUID mID; // 128 bits
S32 mSize; // total size of image if known (NOT size cached)
U32 mTime; // seconds since 1/1/1970
};
</pre>
:*Version 1.2: ''Array of 28 byte entries''
<pre>
struct Entry
{
LLUUID mID; // 16 bytes
S32 mImageSize; // total size of image if known
S32 mBodySize; // size of body file in body cache
U32 mTime; // seconds since 1/1/1970
};
</pre>
* cache/texture.cache
''Array of 600 byte entries (first 600 bytes of .j2c file)''
=== The main "body" cache consists of an entries file and 16 subdirectories ===
* cache/textures/texture.entries ''(Not used with cache version 1.2)''
** This file consists of a an array of EntriesInfo, except mSize is the body size, not the total image size
** When a texture is written to the cache, a new EntriesInfo struct is appended to the file ''regardless of whether or not an entry already exists''.
*** This prevents the disk cache from having to explicitly keep track of which entries have bodies stored in the main cache
** On startup and when the disk cache becomes full, textures.cache is parsed
*** Only the latest entry for each texture is retained.
*** Old entries are purged until 10% of the disk cache is available
''Array of 24 byte entries''
<pre>
struct EntriesInfo
{
LLUUID mID; // 128 bits
S32 mSize; // size of texture body stored on disk
U32 mTime; // unused TODO: create separate structure without this entry
};
</pre>
* cache/textures/[0-f]/uuid
''Body of the texture, i.e. all of the texture minus the header''
== Analysis ==
(so far:)
In the discussion of the texture cache has, it is apparent that there are two separate issues involved. Beyond the view of the texture cache, there is the method to which to determine how long a texture is stored in the cache and there is medium format of the textures being stored. At this time, the format of the medium is still under discussion. However, there is significant data to move to a new system for the method to determine the storage time.
* [http://www.vldb.org/dblp/db/conf/vldb/vldb94-439.html 2Q cache system]. Benefits include that the 2Q method has already been studied by other groups and that it is possible to fit this method into the current design. Only drawback noticed so far is that it appears to carry a specific implementation, but this drawback will need to be confirmed. The 2Q's intended implementation may not easily allow for extra values in the cache system to determine factors of pixel coverage. The drawback is of low impact to the the current design since the significant pixel coverage data is stored in memory. However, in order to further implement the ideas being suggested in the discussion about the medium format and pixel coverage, we could adapt the 2Q system.
* [http://www.almaden.ibm.com/StorageSystems/projects/arc/ ARC] Algorithm designed by IBM Research as an improvement over LRU.  Patent encumbered, however IBM will grant license for open source use only.


== Discussion about improvements ==
== Discussion about improvements ==
Main Article: [[SLDEV Discussion - Texture Cache Plan of Attack]]


The texture cache is a hot topic for optimization. Current ideas/subjects discussed from the mailing list include:
The texture cache is a hot topic for optimization. Current ideas/subjects discussed from the mailing list include:

Latest revision as of 09:51, 19 December 2009

LLTextureCache is a worker thread interface class used for reading and writing texture data to the local disk cache (part of the Image System)

Current design

  • There are two different disk caches used by the Second Life Viewer:
    • The static texture cache
      • Located in the Second Life/skins/textures directory
      • This is a read-only cache which contains common textures shipped with the Viewer
    • The general texture cache
      • Stored in a 'textures' subdirectory in the cache directory, specified in the preferences
      • This is a two tiered read/write cache
        • A list of all files in the cache is stored in texture.entries
        • The header for each image (enough to identify the size and load the first mip level) is stored in texture.cache
        • The body (minus the header) for a limited (by the cache size) number of entries is stored in textures/[0-9a-f]/textureid
  • Requests to read or write to the cache are processed in the priority given
  • The results are given through a Responder class when they are ready

Disk storage format

  • Currently all textures in Second Life are stored in the JPEG2000 compressed format

The header consists of two files

  • cache/texture.entries

8 byte header:

	struct EntriesInfo
	{
		F32 mVersion;
		U32 mEntries;
	};
  • Version 1.0: Array of 24 byte entries
	struct Entry
	{
		LLUUID mID; // 128 bits
		S32 mSize; // total size of image if known (NOT size cached)
		U32 mTime; // seconds since 1/1/1970
	};
  • Version 1.2: Array of 28 byte entries
	struct Entry
	{
		LLUUID mID; // 16 bytes
		S32 mImageSize; // total size of image if known
		S32 mBodySize; // size of body file in body cache
		U32 mTime; // seconds since 1/1/1970
	};
  • cache/texture.cache

Array of 600 byte entries (first 600 bytes of .j2c file)

The main "body" cache consists of an entries file and 16 subdirectories

  • cache/textures/texture.entries (Not used with cache version 1.2)
    • This file consists of a an array of EntriesInfo, except mSize is the body size, not the total image size
    • When a texture is written to the cache, a new EntriesInfo struct is appended to the file regardless of whether or not an entry already exists.
      • This prevents the disk cache from having to explicitly keep track of which entries have bodies stored in the main cache
    • On startup and when the disk cache becomes full, textures.cache is parsed
      • Only the latest entry for each texture is retained.
      • Old entries are purged until 10% of the disk cache is available

Array of 24 byte entries

	struct EntriesInfo
	{
		LLUUID mID; // 128 bits
		S32 mSize; // size of texture body stored on disk
		U32 mTime; // unused TODO: create separate structure without this entry
	};
  • cache/textures/[0-f]/uuid

Body of the texture, i.e. all of the texture minus the header

Analysis

(so far:)

In the discussion of the texture cache has, it is apparent that there are two separate issues involved. Beyond the view of the texture cache, there is the method to which to determine how long a texture is stored in the cache and there is medium format of the textures being stored. At this time, the format of the medium is still under discussion. However, there is significant data to move to a new system for the method to determine the storage time.

  • 2Q cache system. Benefits include that the 2Q method has already been studied by other groups and that it is possible to fit this method into the current design. Only drawback noticed so far is that it appears to carry a specific implementation, but this drawback will need to be confirmed. The 2Q's intended implementation may not easily allow for extra values in the cache system to determine factors of pixel coverage. The drawback is of low impact to the the current design since the significant pixel coverage data is stored in memory. However, in order to further implement the ideas being suggested in the discussion about the medium format and pixel coverage, we could adapt the 2Q system.
  • ARC Algorithm designed by IBM Research as an improvement over LRU. Patent encumbered, however IBM will grant license for open source use only.

Discussion about improvements

Main Article: SLDEV Discussion - Texture Cache Plan of Attack

The texture cache is a hot topic for optimization. Current ideas/subjects discussed from the mailing list include:

  • Multi-level cache for uncompressed and compressed textures
  • Example Cache Levels:
    • Memory
    • Uncompressed Disk
    • Compressed Disk
    • Low Rez Compressed
    • Network
    • A flexible policy architecture would be able to discard, cache, retrieve, and migrate texture data from one level to the next as is appropiate with an external XML file to facilitate fine tuning of policy parameters.
  • Fixed block size cache to store images of any size that is a power of 2
  • Reformatted textures, being changed from jpeg2000 to TGA (with or without RLE), BMP, GIF, or downsampled
  • Storage:
    • Filesystem
    • Database
    • Network
    • Memory mapped
  • Quality of cached images
  • System Requirements
    • Minimal
    • Optimal
    • Surveys
    • Core(s)


Please use the Talk page to continue discussion.

Related articles: Image System, VFS