GitLab

Contributed by "Zhang, Jingwang" <Jingwang.Zhang@emc.com>.

- Update README in scripts dir
- Revert "initial testing script implementation"
  This reverts commit 511a1dc7.

FTB service is not receiving further development,
so disable for now.

This commit removes the protocol buffer dependency
(used only to encode the message being broadcasted over FTB)

Resuming an active state machine might lead to an assertion. So if
it is an active state machine, we need to wait for it to complete
first and then resume it.

This reverts commit dad2e6309603cee63f8743b0199784ed90ec8a0f.

The reverted patch is not the right fix. The problem should
be properly fixed by reverting the patch who changed BMI_test
to BMI_testcontext.

Using BMI_testcontext will break multi-threaded applications.

This reverts commit 055c6757.

Previously, we found that when the memory is full of cached pages
(even they are all clean) the performance will drop significantly.
So we clear all cached pages when the file is closed.

However now we found that the problem is in Linux kernel and it has
been fixed in CentOS 6.2. So the previous patch is unnecessary if
we avoid to use SLES 11.

This reverts commit a62a6eea75041e7418f89e8d3dc6798ab67b254a.

Five counters are added by this commit, here is the details:
- write_sm: Counter on how many write state machine tasks are created.
  It will be increased by one when a new write request arrives.
- write_sm_livetime: The duration from a specific write request is
  created to the time it is deleted.
- write_sm_recvtime: The time duration spent to receive the data buffer
  from the client in a write request.
- write_sm_waittime: The time duration spent in waiting for CPU in the
  task queue for every write request.
- write_sm_fsiotime: The time duration spent in performing IO requests
  to the backend file system.

This commit added a dumper thread to dump all the counters to std::cout
periodically. You can start this thread by giving a positive number to
"counters { dumpinterval = %% }" in the configure file.

These counters could be enabled separately in the configure file, this
commit also demonstrated how to do that.

Change-Id: Ieaf4d1f6c2ed71e8c17d...

The old default thread pool will create lots of short-lived threads
(and was the optimal one for BG/P systems running Zepto);
However, on most other systems, this is not the optimal behaviour,
especially under memory pressure.

Switch default to Cray Threadpool implementation.

Make calls autoconf in some cases and the boggus extra "$" gets reintroduced into the configure script, despite the previous fix.
This will cause g++ to fail by supplying a boggus filename, which in turn will show as an error like "cannot compute the size of size_t".

I have traced it back to this file.
If removed, then one can execute $./configure ... followed by $make without bumping into the issue.

Change-Id: Ie65765c82bbbb7b8fa78303d7663504c3141cbee

Otherwise, the iterator might be invalidated by erase().

Implement zoidfs_commit() call to drop all the page caches belong
to the given file. So that the client can use this call to release
page caches in the server side.

Release the dcache entry and close the file when it is deleted. So
that underlying resources could be released.

context cannot be shared among threads, so we should create context
for every thread.
In the fix, I create a BMI context in the private data of this thread.

Fixes #115

Move each frontend into its own directory

IOFWDMain was treating the frontend as a service,
but the header included was still specific to the BMI service.
Fixed.