File System Performance and Durability

Required reading: Rethink the Sync.
EXT3, Rethink the Sync

this lecture
 finish ext3
 Rethink the Sync

how does ext3 get good performance despite logging entire blocks?
 batches many syscalls per "transaction"
 defers copying cache block to log until it commits log to disk
 hopes multiple sycalls modified same block
   thus many syscalls, but only one copy of block in log

sys call:
 h = start()
 get(h, block #)
   warn logging system we'll modify cached block
     added to list of blocks to be logged
   prevent writing block to disk until after xaction commits
 modify the blocks in the cache
 guarantee: all or none

ext3 transaction
 while "open", adds new syscall handles, and remembers their block #s
 only one open transaction at a time
 ext3 decides to commit current transaction every few seconds (or  

committing a transaction to disk
 mark transaction as done (new syscalls must start new xaction)
 wait for in-progress syscalls to stop() (or not?)
 for all blocks in list to be logged
   append descriptor w/ block # to on-disk log
   append block content from cache to on-disk log
 wait for all log writes to finish
 append the commit record
 (now blocks mentioned in committed log can go to disk)

a new xaction may start while prev is committing
 what if syscall in new xaction wants to change a block in old  
committing xaction?
 can't be allowed to write buffer that old xaction is writing to disk
   then one of new syscall's writes would be visible in old xaction
 new transaction gets a separate copy of the buffer to modify
   old transaction holds onto old copy to write to log
 are there now *two* versions of the block in the buffer cache?
   no, on the new one is in the buffer cache, the old one isn't
   ext3 doesn't write old copy to FS, since new xaction will write it
   but now cannot free up old transaction in log until new one commits
     don't want to lose only copy of old xaction's committed updates

how to free up log space
 after commit, cached blocks written to real locations in FS on disk
 then that part of log can be re-used

what if a crash?
 crash may interrupt writing the log -> partial transaction in log
 may not have finished writing cache to disk
   only relevant for a committed transaction

how does recovery work
 1. find the start and end of the log
    superblock start and seq# of first transaction
    scan until bad record or not the expected seq #
    go back to last commit record
    crash during commit -> last transaction ignored during recovery
 2. replay all blocks from whole transactions, in log order

other interesting tidbits about ext3:

ext3 not as durable as xv6
 creat() returns -> maybe data is not on disk! crash will undo it.
 need fsync(fd) to force commit of current transaction, and wait
 would ext3 have good performance if commit after every sys call?
   would log many more blocks, no absorption
   10 ms per syscall, rather than 0 ms
   xsyncfs Figure 4 shows 10x decrease in performance!
 (Rethink the Sync addresses this problem)

lack of checksum in ext3 commit record
 disks usually have write caches and re-order writes, for performance
   sometimes hard to turn off (the disk lies)
   people often leave re-ordering enabled for speed, out of ignorance
 bad news if disk re-orders writes, writes commit before preceding  
 then recovery replays descriptors with random block #s!
 and writes them with random content!

ordered vs journaled
 journaling file content is slow, every data block written twice
 in principle not needed to keep FS internally consistent
 but can't just write data blocks whever we feel like it:
   if metadata updated first, crash may leave file pointing
   to blocks with someone else's data
 ext3 ordered mode:
   write() causes data blocks to go to disk before committing adding  
block # to inode
   thus won't see stale data if there's a crash

correctness challenges w/ ordered:
 A. rmdir, re-use block for file, ordered write of file,
      crash before rmdir or write committed
    now scribbled over the directory block
    fix: defer free of block until freeing operation forced to log  
on disk
 B. rmdir, commit, re-use block in file, ordered file write, log  
      crash, replay rmdir
    file is left w/ directory content e.g. . and ..
    fix: revoke records, prevent log replay of a given block

Rethink the Sync

intro: why are we reading Rethink the Sync?
 ext3 defends FS internal consistency (no dirent -> unallocated  
inode &c)
 but that is not enough!
   apps and external users also need consistency
   ext3 may hurt here: delayed commit

xsyncfs's promises to external user
 correctness of synchronous FS updates
 performance of asynchronous FS

what is a sync FS?
 a sync FS forces updates to disk before returning from syscall,  
like xv6
   if syscall returns, its effects will be visible after a crash
 write-through disk cache in kernel, so slow

why is sync FS a good standard for correct FS behavior?
 maybe since sync FS easy to reason about?
   can count on all ops completed before crash being visible after  
   e.g. write data to file, print "OK"
 no-one uses a sync FS or writes apps that assume one...
 in real life: write(), fsync(), print "OK"

what does paper mean by an async FS?
 system calls can return before data is on disk
   syscalls modify blocks in a write-back cache
   but then FS has to worry about internal consistency
 paper really means FS w/ async logging, in particular Linux's ext3

why is not-durable a problem for external users?
 mail server says "OK" but crash erases its copy of msg
   OK -> sending server deleted, so now msg entirely gone
   % cp * backupdir
   is it ok now to edit the original files?

currently you have to choose between durability and performance
 slow sync fs, or fast non-durable async logging
 or remember to call fsync(), maybe call it too much

what's the paper's insight to get performance + durability?
 sufficient definition of durability:
    updates only need to be durable by the time of external output
    display output, network packet
 if user sees no output, no reason to expect FS I/O has finished

which data has to be durable before external output?
 certainly all writes &c of process producing the output
 also all updates "causally preceding" the external output
   X : write(file) -> exit()
   Y :                       -> wait() -> printf("$ ")
   process X's file write must be on disk before user sees Y's "$ "

why is durability of causally preceding updates useful to external  
 if mail sender sees "OK" msg via net, guaranteed msg is on rcvr's  
   no "OK" -> recovery by resending later
 % cp * backupdir
 if you see the 2nd prompt, you know backup is complete
 if crash before 2nd prompt, user "recovers" by backing up again
 seems to be sufficient in these two cases

can we argue that no non-causally-preceding write is necessary?
   read mail msg from socket
   fork a child
     child writes msg to a file
   parent writes "OK" to socket
 so: sufficient only if application is written carefully

is it necessary for xsyncfs to force all causally preceding writes?
 examples when maybe app doesn't need a causally preceding write  

will xsyncfs have higher performance than sync updates?
 can use write-back cache, coalesce writes
 external output forces batched disk update
 will be fast if many updates to same disk block

will xsyncfs have higher performance than async logging?
 usually no:
   xsyncfs has causal tracking overhead
   xsyncfs might force to disk more often than every 5 seconds

when might xsyncfs have *higher* performance than async FS?
 mysql example
   client does multiple SQL xactions before external output
   mysql server must fsync() after each xaction for D in ACID
   but xsyncfs knows: no external output -> can ignore fsync()

what guarantees does xsyncfs give to applications?
 paper implies apps written for sync file system will work fine
   xsyncfs guarantees ordering: section 2.2, end of page
   so like sync FS, but crashed earlier
   your app still needs to be able to recover from a crash
     at any point, can rely on all syscalls up to a certain
     point being visible after the crash
     examples of why prefix property is useful in app recovery?

what are the hard parts of xsyncfs design / impl?
 tracking causality
   each process and kernel object (pipe) has list of not-yet-written  
   copy that list when a process reads/updates kernel object
 buffering external output, e.g. display writes and network packets
   why useful to buffer? why not just suspend process?
 forcing disk writes on external output
   track down all causally preceding blocks
   ask ext3 to write them
   wait, then release output
   what does that involve?
     can you ask ext3 to write just a particular block?
     no: xsyncfs must force whole log, including unrelated updates

given that ext3 won't let you write an individual block
 why do they have to do all that causal tracking?
 why not just commit ext3 xaction when program does external output?

this eliminates a potential advantage of xsyncfs over async logging FS
 in principle, if there are multiple unrelated apps,
   xsyncfs only needs to force causally related updates
   not entire log, which is what ordinary ext3 fsync() does
 in practice, xsyncfs forces whole ext3 log

xsyncfs evaluation
 what do they need to demonstrate?
   that the problem they're solving exists
   that they achieve the main goals stated at the start of the paper
 #1. synchronous file systems are slow (i.e. problem exists)
 #2. asynchronous file systems aren't very durable (ditto)
 #3. xsyncfs as durable as a synchronous file system
 #4. xsyncfs as fast as an asynchronous file system

figure 3: durability
 (show #2 and #3)
 very cool that they systematically test crashes
 async is ordinary ext3, force every 5 seconds or on fsync
 sync is force ext3 log after every system call
 sync + write barriers fixes ide write caching
 xsyncfs uses ext3 and write barriers
 comparison unfair: async ext3 should be using write barriers
   that is how it was designed to be used
   then it would be as durable (and as high performance) as xsyncfs

figure 4:
 single-threaded read/write/create/delete benchmark
 no external output
 lots of FS I/O
 would we expect xsyncfs to be faster than async ext3?
   why or why not?
   shows #4
 why is sync ext3 so slow?
   shows #1

figure 6: mysql, vs async ext3 with write barriers
 mysql fsync()s log after each SQL transaction
   so both systems are arguably equally durable
 why does xsyncfs win with few threads?
 why does ext3 catch up with many threads?

figure 7: web server, not much difference, why?
 what does specweb99 do?

what is the bottom line?
 should linux switch from ext3 to xsyncfs?