lock.  The lock is now never held for a long time, only while
self.thread is being assigned or inspected, and while a thread is
being created, started or stopped.  Waiting for an event or joining a
thread is now done without holding the lock.  There is no longer a
fear of AttributeError.

When a thread join doesn't succeed within 30 seconds, a message is
logged, and depending on the situation, the join is retried.

The old code would avoid hosing the system when the first jar failed
during its finish.  The logic is that the failure means that
transaction did not commit, so it was safe for all the other jars to
abort.  There are counter-examples to this assumption.  For example,
ZEO could fail after commiting the transaction on the server but
before communicating that success to the client.  It would lead to
inconsistency if the other jars aborted after the first committed.

The solution is to reduce the special case yet further:  If a
single-jar transaction fails during the finish, assume the transaction
is aborted and avoid hosage.  If the jar actually committed, it's
possible to get back to a consistent state since only one jar was
involved.

The effect of this change will be to cause transactions to fail
immediately, rather than the first time they call store() or
abortVersion() or some other write-method.  Since the Connection now
treats empty transactions as a special case, this change will not
change the class of transactions that fail.

parenthesis.

test suite.  The address is now found as the _addr attribute of the
client storage instance.

the BuddyCacheSimulation constructor to round up the cachelimit so
that it reports the actual value used.

to print some information.  Use -s to print size histograms.  This is
pretty raw.  It prints two tables: one for writes, one for loads.

keeping track of the beginning and end address of all free blocks, to
make merging of adjacent blocks a breeze.  The free list is no longer
kept sorted order.

Unfortunately, this still performs poorly.  It has a higher hit rate
(with the 38 minute trace I'm using; the longer traces take too long)
than the buddy system allocator with the same arena size, it is *much*
slower, the memory usage is under 20% (!!!), and despite a roving
pointer, the allocation loop is taken an average of 692 times per
allocation.  At the end of the simulation the free list contains 1039
blocks; I presume this is a steady state approximating the average
behavior.  (There are 2280 allocated blocks at that point, roughly
confirming Knuth's "50% rule" and suggesting that the block merge code
works properly.)

I guess the next idea is separate free lists per size.

tpc_begin(): Get rid of the apply() call since I believe we can now
rely on extended-call syntax being available.

simple free list using a mix of exact-fit and last-fit with a roving
pointer as suggested by Tim.  Use -f to use this version.

Renamed BuddyNode to BlockNode since it is now used by both allocators.

Moved some code around in an attempt to organize it more top-down.

Moved the report printing into the Simulation base class.  Print the
class name and exact cache size before the header; skip the final
report printing if only one previous report was printed.

  Versions:   0 records used a version

1000*1000.  This makes it easier to specify cache sizes for the buddy
system that don't have to be rounded up to the next power of two.  It
does make future reported results incompatible with past results.

allocator.

Different versions of ZODB do different things with the serialnos for
abortVersion().  To be safe and avoid giving non-version data with an
invalid serialno after an abort version, invalidate the version and
non-version data.

This hurts cache effectiveness when a version is aborted, but I expect
it will have little practical impact.

Sort options alphabetically.  Remove dependency between -S, -q and -v
options (-S no longer disables -q and enables -v).  Remove dependency
between -S and -h (histogram is printed regardless of dostats).

in, right, Jeremy? :-)

Repair comments in _call() about how wait() handles reply lock.

# NB: commit() is responsible for calling tpc_begin() on the storage.
# It uses self._begun to track whether it has been called.  When
# self._begun is None, it has not been called.

# This arrangement allows us to handle the special case of a
# transaction with no modified objects.  It is possible for
# registration to be occur unintentionally and for a persistent
# object to compensate by making itself as unchanged.  When this
# happens, it's possible to commit a transaction with no modified
# objects.

# Since tpc_begin() may raise a ReadOnlyError, don't call it if there
# are no objects.  This avoids spurious (?) errors when working with
# a read-only storage.

Add code to handle this in Connection's tpc_begin() and commit()
methods.

Add two tests in testZODB.

An earlier revision (1.68) fixed invalidation code when an object had
no oid (oid == None).  There is no longer anyway to get None in
_invalidated, and it has been a long time since a None there meant
"invalidate everything."

At long last, then, remove all the code deal with the invalidate
everything behavior.

  invalidations instead of only the ones that hit the cache.

- Support reading gzipped files.  This is triggered automatically when
  the filename ends in .gz.  Also support reading from stdin (pass '-'
  as filename).

- Widen the FLIPS column by one to match the EVICTS column width of
  the LRU simulation.

- Support simulating alternative caching strategies.  For now, only
  LRU is supported, triggered by the -l option.

- Change the Simulation base class to store the cachelimit argument.
  (The constructor signature must be the same across all subclasses.)

the filename ends in .gz.  Also support reading from stdin (pass '-'
as filename).

base class.

report overall statistics for all columns.

ZEO_CACHE_TRACE set (see stats.py for file format description).

correct header sizes.