[SSI-users] OpenSSI 1.1.0 released for Fedora Core 2

[Brian J. W. <Bri...@hp...>]

The OpenSSI 1.1.0 development release for Fedora Core 2 is now available 
from OpenSSI.org. The versions for Debian and Red Hat 9 will follow in 
several days.

The release notes are below.

Enjoy!

Brian

---
OpenSSI 1.1.0 is a development release. If you want a stable release
that is more suitable for production use, please download OpenSSI 1.0.0
for Red Hat 9 or Debian. If you are more interested in having the latest
features, then read on!!

This is the first release to have a version for Fedora Core 2 ("FC2").
It also has versions for Debian and Red Hat 9 ("RH9"). The RH9 version is
only recommended for users upgrading from OpenSSI 1.0.0 for RH9. All other
users should be on FC2 or Debian going forward.

The OpenSSI kernel is now based on the most recent Fedora Core 1 ("FC1")
kernel (2.4.22-1.2199.nptl). This is true for all distributions, since it
is difficult to maintain multiple versions of the OpenSSI kernel. It might
seem strange that OpenSSI 1.1.0 runs on FC2, but it is based on the FC1
kernel.  This is because the FC1 kernel is based on the Linux 2.4 kernel,
whereas FC2 is based on 2.6. There is currently a project to port OpenSSI
to the 2.6 kernel, but it will not be ready for at least two months. When
it is ready, it will be part of OpenSSI 1.9.0, while the OpenSSI 1.1.x
series (based on the FC1 kernel) stabilizes into OpenSSI 1.2.0 for
production use.

There are several features that have been developed for the 1.1.0 release
during the several months that 1.0.0 was stabilizing. One of them is
performance enhancements for the Cluster File System ("CFS"). Now it not
only caches remote reads, but it also caches remote writes, while still
maintaining a coherent view of the filesystem across the cluster.
Furthermore, CFS now does asynchronous remote read-aheads of data blocks
that programs might want, so that they are already cached locally by time
the program wants them. Hopefully you will notice a significant performance
improvement in your filesystem intensive application!

Another new feature is atomic migration of a group of processes. The
group could either be a POSIX process group or the "threads" of a multi-
threaded application (on Linux, each thread is a full process). To
migrate a POSIX process group, call the migrate command with the negative
PID of the process group leader. These semantics are very similar to
signaling a process group. To migrate a thread group, call the migrate
command with the PID of any "thread" in that group. With both forms of
group migration, either the entire group migrates or it does not. If any
process in the group is unable to migrate for any reason, the entire
group will remain on the old node.

Another process migration enhancement is the ability to migrate a process
while it is holding file record locks. These locks will continue to be
held during the migration and after the process continues running on the
new node.

LVS-NAT can now be used with OpenSSI. Linux Virtual Server ("LVS") is a
third-party open-source project that load balances TCP connections among
the nodes in a cluster. For a long time, LVS has been integrated with
OpenSSI, making it easier to manage than an LVS cluster without OpenSSI,
but only the Direct Routing ("DR") feature of LVS was supported. DR
allows load balancing in a cluster where every node has a direct
connection to the external network (as well as a network connection to a
private switch for the cluster interconnect, as recommended in the
OpenSSI installation instructions). Unfortunately, a security feature in
the Linux kernel prevents DR from being used in a cluster where only some
nodes are connected to the external network. For these situations, the
Network Address Translation ("NAT") feature of LVS should be used, and it
is now supported on OpenSSI.

Note that LVS-NAT is different from the NAT that you would use for making
outbound connections from a private IP address. LVS-NAT is for load
balancing inbound connections from a public IP address among a cluster
of machines that are all connected to a private network, such as an
OpenSSI cluster interconnect. Of course, LVS-NAT requires that potential
director nodes are connected to both the external network and the cluster
interconnect, so that traffic can move between the two networks.

The 'fast' and 'fastnode' commands have been added. 'fastnode' returns the
node number for the least-loaded node in the cluster, as determined by the
process load-leveling algorithm. 'fast' executes a command on the least-
loaded node. Read the man pages for these commands for more information.

Several files were added to /proc/cluster/: nm_rate, nm_log_threshold,
and nm_nodedown_disabled. nm_rate can be used to alter how often node
monitoring messages are exchanged (default is 1 per second) and how long
before a node is declared down (default 10 seconds). nm_log_threshold
indicates how many monitoring messages can be missed before a kernel
warning is generated (default 2). nm_nodedown_disabled can be set to
disable nodedown detection, which is useful if you need to enter the
kernel debugger on one of the nodes. Previously, you had to recompile
an OpenSSI kernel to change one these values. Now you can do it by
simply writing a new value into one of these /proc/cluster files.

A variety of bugs have been fixed. For more information, see the
ChangeLog.