The status of the Squid cyclic fs (COSS) :

COSS was originally implemented as an on-disk LRU. I'll describe the original implementation as I grabbed from Eric Stern now.

A filesystem is just a single large file or physical device.

A membuf - 1 megabyte in size - is initially allocated to represent the first megabyte of the filesystem. Objects are copied into the membuf if their size is known up front (and thus space can be 'pre-allocated' in the stripe.) When the stripe is filled up it is marked as "full" and written to the filesystem. Objects are added to the beginning of a linked list as this happens.

Objects are referenced by their offset on the disk: any read is first checked against the in-memory membuf list. If an object is found to be in a membuf then a copy of the object data is taken and the data is handed back to Squid. If an object is not found in a membuf it is read from the filesystem, placed at the head of the current membuf - and they are re-added to the head of the linked list - and the squid file pointer is updated to point to this new position.

As stripes are successively allocated and written to the filesystem in order the 'popular' objects stay near the 'head'. This happens until the last stripe on disk is written: at which point the write pointer is cycled to the beginning of the filesystem.

At this point the LRU implementation kicks in: the objects which are at the end of this linked list match those at the beginning of the filesystem. COSS will start at the end of the linked list and move backwards, deallocating objects, until it reaches the beginning of the next stripe. It then has enough room to allocate a 1 megabyte stripe (and its membuf.) at the beginning of the disk. It then fills this membuf as described above.

When this membuf is filled it writes the stripe to disk, frees the objects in the next megabyte of disk and then allocates a membuf and fills that.

This implementation wasn't complete:

• The rebuild-from-logfile didn't seem to work right
• There was no support for rebuild-from-disk (in case the logfile was missing or corrupt)
• The implementation used file_read() and file_write() - callback methods of scheduling disk filedescriptor IO - but assumed synchronous behaviour.

When I adapted the codebase to use POSIX AIO I discovered a number of race conditions in the COSS code:

• Objects which were being read from disk and written into the current membuf had their squid file pointer numbers updated. Subsequent reads of this object would be copied from the current membuf - but async disk IO wouldn't guarantee the data there was written until some time after scheduling. This resulted in a lot of swapin failures as NULL data was being written
• It was possible, but so far unlikely, that a disk operation would be scheduled for an object which was then overwritten by a full stripe.

The nice features of COSS was the simple writing and object pool maintainence: writes were aggregated and predictable (being in 1 megabyte chunks.) Popular objects had more of a possibility of staying in the current membuf.

I recently took the code and began fixing the bugs. These included:

• All disk stripes were now an even multiple of the membuf size (1 megabyte.) Eric's original implementation would note when a membuf was free, write the membuf to disk and then start the new membuf at the end of the old membuf. This meant a few bytes weren't being wasted but it did make dividing the filesystem up for analysis/repair/rebuild difficult.
• Object relocations are now tracked from their creation to completion
• When an object is relocated its data - and any subsequent read request - is stalled until the object data has been read in from the filesystem.
• A check (and loud log message!) has been added to catch attempts to write a stripe where a pending relocate is occuring (and the read hasn't completed), hopefully catching (but not repairing for now) instances where said read will result in then-bogus data
• Rebuild logic has been added - its now easy to read the disk in as 1 megabyte chunks and do basic checks on each stripe. If a stripe has been partially or badly written to disk the contents can be thrown away without affecting the rest of the filesystem
• Objects no longer live in a single linked list. Each on-disk stripe reigon has an in-memory structure used to track various statistics including a linked list containing which objects are currently there. This makes freeing any arbitrary stripe easy, allowing for much cleaner object expiry and filesystem index rebuild process.

The problems seen so far:

• The write rate is a function of not only the cacheable data coming in from the network but the hit rate - and subsequent relocation of popular objects - which means the write volume can quickly spiral out of control
• Some hit-rate issues which I haven't figured out yet. It may be related to my relatively small test caches (~ 8-10 gigabytes) and the polygraph workloads using a much bigger cache data set.

Possible directions to take (although I do need some actual world-testing and statistics first!):

• Find out what percentage of objects are read in and never referenced again vs objects re-referenced once, twice, four times, eight times, etc.
• Consider allocating stripe areas as "hot object" stripes which aren't part of the disk LRU. Place popular objects in these stripes and don't relocate them once they're there - this should cut down on the constant object relocation and therefore cut back on the write bandwidth. They can be managed by a different LRU or other replacement scheme.
• Consider implementing some form of object locality; which will need cooperation from other areas of squid.

Interested in the work? I'm placing snapshots up on my squid website - here.