path: root/docs/howto/HighAvailability.xml

<chapter id="SambaHA">
<chapterinfo>
	&author.jht;
	&author.jeremy;
</chapterinfo>

<title>High Availability</title>

<sect1>
<title>Features and Benefits</title>

<para>
Network administrators are often concerned about the availability of file and print
services. Network users are inclined toward intolerance of the services they depend
on to perform vital task responsibilities.
</para>

<para>
A sign in a computer room served to remind staff of their responsibilities. It read:
</para>

<blockquote>
<para>
All humans fail, in both great and small ways we fail continually. Machines fail too.
Computers are machines that are managed by humans, the fallout from failure
can be spectacular. Your responsibility is to deal with failure, to anticipate it
and to eliminate it as far as is humanly and economically wise to achieve.
Are your actions part of the problem or part of the solution?
</para>
</blockquote>

<para>
If we are to deal with failure in a planned and productive manner, then first we must
understand the problem. That is the purpose of this chapter.
</para>

<para>
Parenthetically, in the following discussion there are seeds of information on how to
provision a network infrastructure against failure. Our purpose here is not to provide
a lengthy dissertation on the subject of high availability. Additionally, we have made
a conscious decision to not provide detailed working examples of high availability
solutions; instead we present an overview of the issues in the hope that someone will
rise to the challenge of providing a detailed document that is focused purely on
presentation of the current state of knowledge and practice in high availability as it
applies to the deployment of Samba and other CIFS/SMB technologies.
</para>

</sect1>

<sect1>
<title>Technical Discussion</title>

<para>
The following summary was part of a presentation by Jeremy Allison at the SambaXP 2003
conference that was held at Goettingen, Germany, in April 2003. Material has been added
from other sources, but it was Jeremy who inspired the structure that follows.
</para>

	<sect2>
	<title>The Ultimate Goal</title>

	<para>
	All clustering technologies aim to achieve one or more of the following:
	</para>

	<itemizedlist>
		<listitem><para>Obtain the maximum affordable computational power.</para></listitem>
		<listitem><para>Obtain faster program execution.</para></listitem>
		<listitem><para>Deliver unstoppable services.</para></listitem>
		<listitem><para>Avert points of failure.</para></listitem>
		<listitem><para>Exact most effective utilization of resources.</para></listitem>
	</itemizedlist>

	<para>
	A clustered file server ideally has the following properties:
	</para>

	<itemizedlist>
		<listitem><para>All clients can connect transparently to any server.</para></listitem>
		<listitem><para>A server can fail and clients are transparently reconnected to another server.</para></listitem>
		<listitem><para>All servers server out the same set of files.</para></listitem>
		<listitem><para>All file changes are immediately seen on all servers.</para>
			<itemizedlist><listitem><para>Requires a distributed file system.</para></listitem></itemizedlist></listitem>
		<listitem><para>Infinite ability to scale by adding more servers or disks.</para></listitem>
	</itemizedlist>

	</sect2>

	<sect2>
	<title>Why Is This So Hard?</title>

	<para>
	In short, the problem is one of <emphasis>state</emphasis>.
	</para>

	<itemizedlist>
		<listitem>
			<para>
			All TCP/IP connections are dependent on state information.
			</para>
			<para>
			The TCP connection involves a packet sequence number. This
			sequence number would need to be dynamically updated on all
			machines in the cluster to effect seamless TCP fail-over.
			</para>
		</listitem>
		<listitem>
			<para>
			CIFS/SMB (the Windows networking protocols) uses TCP connections.
			</para>
			<para>
			This means that from a basic design perspective, fail-over is not
			seriously considered.
			<itemizedlist>
				<listitem><para>
				All current SMB clusters are fail-over solutions
				&smbmdash; they rely on the clients to reconnect. They provide server
				fail-over, but clients can lose information due to a server failure.
				</para></listitem>
			</itemizedlist>
			</para>
		</listitem>
		<listitem>
			<para>
			Servers keep state information about client connections.
			<itemizedlist>
				<listitem><para>CIFS/SMB involves a lot of state.</para></listitem>
				<listitem><para>Every file open must be compared with other file opens
						to check share modes.</para></listitem>
			</itemizedlist>
			</para>
		</listitem>
	</itemizedlist>

		<sect3>
		<title>The Front-End Challenge</title>

		<para>
		To make it possible for a cluster of file servers to appear as a single server that has one
		name and one IP address, the incoming TCP data streams from clients must be processed by the
		front end virtual server. This server must de-multiplex the incoming packets at the SMB protocol
		layer level and then feed the SMB packet to different servers in the cluster.
		</para>

		<para>
		One could split all IPC$ connections and RPC calls to one server to handle printing and user
		lookup requirements. RPC Printing handles are shared between different IPC4 sessions &smbmdash; it is
		hard to split this across clustered servers!
		</para>

		<para>
		Conceptually speaking, all other servers would then provide only file services. This is a simpler
		problem to concentrate on.
		</para>

		</sect3>

		<sect3>
		<title>De-multiplexing SMB Requests</title>

		<para>
		De-multiplexing of SMB requests requires knowledge of SMB state information,
		all of which must be held by the front-end <emphasis>virtual</emphasis> server.
		This is a perplexing and complicated problem to solve.
		</para>

		<para>
		Windows XP and later have changed semantics so state information (vuid, tid, fid)
		must match for a successful operation. This makes things simpler than before and is a
		positive step forward.
		</para>

		<para>
		SMB requests are sent by vuid to their associated server. No code exists today to
		affect this solution. This problem is conceptually similar to the problem of
		correctly handling requests from multiple requests from Windows 2000
		Terminal Server in Samba.
		</para>

		<para>
		One possibility is to start by exposing the server pool to clients directly.
		This could eliminate the de-multiplexing step.
		</para>

		</sect3>

		<sect3>
		<title>The Distributed File System Challenge</title>

		<para>
<indexterm><primary>Distributed File Systems</primary></indexterm>
		There exists many distributed file systems for UNIX and Linux.
		</para>

		<para>
		Many could be adopted to backend our cluster, so long as awareness of SMB
		semantics is kept in mind (share modes, locking and oplock issues in particular).
		Common free distributed file systems include:
<indexterm><primary>NFS</primary></indexterm>
<indexterm><primary>AFS</primary></indexterm>
<indexterm><primary>OpenGFS</primary></indexterm>
<indexterm><primary>Lustre</primary></indexterm>
		</para>

		<itemizedlist>
			<listitem><para>NFS</para></listitem>
			<listitem><para>AFS</para></listitem>
			<listitem><para>OpenGFS</para></listitem>
			<listitem><para>Lustre</para></listitem>
		</itemizedlist>

		<para>
		The server pool (cluster) can use any distributed file system backend if all SMB
		semantics are performed within this pool.
		</para>

		</sect3>

		<sect3>
		<title>Restrictive Constraints on Distributed File Systems</title>

		<para>
		Where a clustered server provides purely SMB services, oplock handling
		may be done within the server pool without imposing a need for this to
		be passed to the backend file system pool.
		</para>

		<para>
		On the other hand, where the server pool also provides NFS or other file services,
		it will be essential that the implementation be oplock aware so it can
		interoperate with SMB services. This is a significant challenge today. A failure
		to provide this will result in a significant loss of performance that will be
		sorely noted by users of Microsoft Windows clients.
		</para>

		<para>
		Last, all state information must be shared across the server pool.
		</para>

		</sect3>

		<sect3>
		<title>Server Pool Communications</title>

		<para>
		Most backend file systems support POSIX file semantics. This makes it difficult
		to push SMB semantics back into the file system. POSIX locks have different properties
		and semantics from SMB locks.
		</para>

		<para>
		All <command>smbd</command> processes in the server pool must of necessity communicate
		very quickly. For this, the current <parameter>tdb</parameter> file structure that Samba
		uses is not suitable for use across a network. Clustered <command>smbd</command>'s must use something else.
		</para>

		</sect3>

		<sect3>
		<title>Server Pool Communications Demands</title>

		<para>
		High speed inter-server communications in the server pool is a design prerequisite
		for a fully functional system. Possibilities for this include:
		</para>

		<itemizedlist>
			<listitem><para>
			Proprietary shared memory bus (example: Myrinet or SCI [Scalable Coherent Interface]).
			These are high cost items.
			</para></listitem>
		
			<listitem><para>
			Gigabit ethernet (now quite affordable).
			</para></listitem>
		
			<listitem><para>
			Raw ethernet framing (to bypass TCP and UDP overheads).
			</para></listitem>
		</itemizedlist>

		<para>
		We have yet to identify metrics for  performance demands to enable this to happen
		effectively.
		</para>

		</sect3>

		<sect3>
		<title>Required Modifications to Samba</title>

		<para>
		Samba needs to be significantly modified to work with a high-speed server inter-connect
		system to permit transparent fail-over clustering.
		</para>

		<para>
		Particular functions inside Samba that will be affected include:
		</para>

		<itemizedlist>
			<listitem><para>
			The locking database, oplock notifications,
			and the share mode database.
			</para></listitem>

			<listitem><para>
			Failure semantics need to be defined. Samba behaves the same way as Windows.
			When oplock messages fail, a file open request is allowed, but this is 
			potentially dangerous in a clustered environment. So how should inter-server
			pool failure semantics function and how should this be implemented?
			</para></listitem>

			<listitem><para>
			Should this be implemented using a point-to-point lock manager, or can this
			be done using multicast techniques?
			</para></listitem>

		</itemizedlist>

		</sect3>
	</sect2>

	<sect2>
	<title>A Simple Solution</title>

	<para>
	Allowing fail-over servers to handle different functions within the exported file system
	removes the problem of requiring a distributed locking protocol.
	</para>

	<para>
	If only one server is active in a pair, the need for high speed server interconnect is avoided.
	This allows the use of existing high availability solutions, instead of inventing a new one.
	This simpler solution comes at a price &smbmdash; the cost of which is the need to manage a more
	complex file name space. Since there is now not a single file system, administrators
	must remember where all services are located &smbmdash; a complexity not easily dealt with.
	</para>

	<para>
	The <emphasis>virtual server</emphasis> is still needed to redirect requests to backend
	servers. Backend file space integrity is the responsibility of the administrator.
	</para>

	</sect2>

	<sect2>
	<title>High Availability Server Products</title>

	<para>
	Fail-over servers must communicate in order to handle resource fail-over. This is essential
	for high availability services. The use of a dedicated heartbeat is a common technique to
	introduce some intelligence into the fail-over process. This is often done over a dedicated
	link (LAN or serial).
	</para>

	<para>
<indexterm><primary>SCSI</primary></indexterm>
	Many fail-over solutions (like Red Hat Cluster Manager, as well as Microsoft Wolfpack)
	can use a shared SCSI of Fiber Channel disk storage array for fail-over communication.
	Information regarding Red Hat high availability solutions for Samba may be obtained from:
	<ulink url="http://www.redhat.com/docs/manuals/enterprise/RHEL-AS-2.1-Manual/cluster-manager/s1-service-samba.html">www.redhat.com.</ulink>
	</para>

	<para>
	The Linux High Availability project is a resource worthy of consultation if your desire is
	to build a highly available Samba file server solution. Please consult the home page at
	<ulink url="http://www.linux-ha.org/">www.linux-ha.org/.</ulink>
	</para>

	<para>
	Front-end server complexity remains a challenge for high availability as it needs to deal
	gracefully with backend failures, while at the same time it needs to provide continuity of service
	to all network clients.
	</para>
	
	</sect2>

	<sect2>
	<title>MS-DFS: The Poor Man's Cluster</title>

	<para>
<indexterm><primary>MS-DFS</primary></indexterm>
<indexterm><primary>DFS</primary><see>MS-DFS, Distributed File Systems</see></indexterm>
	MS-DFS links can be used to redirect clients to disparate backend servers. This pushes
	complexity back to the network client, something already included by Microsoft.
	MS-DFS creates the illusion of a simple, continuous file system name space, that even
	works at the file level.
	</para>

	<para>
	Above all, at the cost of complexity of management, a distributed (pseudo-cluster) can
	be created using existing Samba functionality.
	</para>

	</sect2>

	<sect2>
	<title>Conclusions</title>

	<itemizedlist>
		<listitem><para>Transparent SMB clustering is hard to do!</para></listitem>
		<listitem><para>Client fail-over is the best we can do today.</para></listitem>
		<listitem><para>Much more work is needed before a practical and manageable high
				availability transparent cluster solution will be possible.</para></listitem>
		<listitem><para>MS-DFS can be used to create the illusion of a single transparent cluster.</para></listitem>
	</itemizedlist>

	</sect2>

</sect1>
</chapter>