Diskless node

With essentially a single OS image for an array of machines (with perhaps some customizations for differences in hardware configurations among the nodes), installing software and maintaining installed software can be more efficient. Furthermore, any system changes made during operation (due to user action, worms, viruses, etc.) can be either wiped out when the power is removed (if the image is copied to a local RAM disk) such as Windows XP Embedded remote boot[9][10] or prohibited entirely (if the image is a network filesystem). This allows use in public access areas (such as libraries) or in schools etc., where users might wish to experiment or attempt to "hack" the system.

However, it is not necessary to implement network booting to achieve either of the above advantages - ordinary PCs (with the help of appropriate software) can be configured to download and reinstall their operating systems on (e.g.) a nightly basis, with extra work compared to using shared disk image that diskless nodes boot off.

Modern diskless nodes can share the very same disk image, using a 1:N relationship (1 disk image used simultaneously by N diskless nodes). This makes it very easy to install and maintain software applications: The administrator needs to install or maintain the application only once, and the clients can get the new application as soon as they boot off the updated image. Disk image sharing is made possible because they use the write cache: No client competes for any writing in a shared disk image, because each client writes to its own cache.

All the modern diskless nodes systems can also use a 1:1 Client-to-DiskImage relationship, where one client "owns" one disk image and writes directly into said disk image. No write cache is used then.

Making a modification in a shared disk image is usually made this way:

1. The administrator makes a copy of the shared disk image that he/she wants to update (this can be done easily because the disk image file is opened only for reading)
2. The administrator boots a diskless node in 1:1 mode (unshared mode) from the copy of the disk image he/she just made
3. The administrator makes any modification to the disk image (for instance install a new software application, apply patches or hotfixes)
4. The administrator shutdowns the diskless node that was using the disk image in 1:1 mode
5. The administrator shares the modified disk image
6. The diskless nodes use the shared disk image (1:N) as soon as they are rebooted.

The use of central disk storage also makes more efficient use of disk storage. This can cut storage costs, freeing up capital to invest in more reliable, modern storage technologies, such as RAID arrays which support redundant operation, and storage area networks which allow hot-adding of storage without any interruption. Further, it means that losses of disk drives to mechanical or electrical failure—which are statistically highly probable events over a timeframe of years, with a large number of disks involved—are often both less likely to happen (because there are typically fewer disk drives that can fail) and less likely to cause interruption (because they would likely be part of RAID arrays). This also means that the nodes themselves are less likely to have hardware failures than fat clients.

Diskless nodes share these advantages with thin clients.

However, this storage efficiency can come at a price. As often happens in computing, increased storage efficiency sometimes comes at the price of decreased performance.

Large numbers of nodes making demands on the same server simultaneously can slow down everyone's experience. However, this can be mitigated by installing large amounts of RAM on the server (which speeds up read operations by improving caching performance), by adding more servers (which distributes the I/O workload), or by adding more disks to a RAID array (which distributes the physical I/O workload). In any case this is also a problem which can affect any client-server network to some extent, since, of course, fat clients also use servers to store user data.

Indeed, user data may be much more significant in size and may be accessed far more frequently than operating systems and programs in some environments, so moving to a diskless model will not necessarily cause a noticeable degradation in performance.

Greater network bandwidth (i.e. capacity) will also be used in a diskless model, compared to a fat client model. This does not necessarily mean that a higher capacity network infrastructure will need to be installed—it could simply mean that a higher proportion of the existing network capacity will be used.

Finally, the combination of network data transfer latencies (physically transferring the data over the network) and contention latencies (waiting for the server to process other nodes' requests before yours) can lead to an unacceptable degradation in performance compared to using local drives, depending on the nature of the application and the capacity of the network infrastructure and the server.

Another example of a situation where a diskless node would be useful is in a possibly hazardous environment where computers are likely to be damaged or destroyed, thus making the need for inexpensive nodes, and minimal hardware a benefit. Again, thin clients can also be used here.

Diskless machines may also consume little power and make little noise, which implies potential environmental benefits and makes them ideal for some computer cluster applications.

• Distributed load The processing load of diskless nodes is distributed. Each user gets its own processing isolated environment, barely affecting other users in the network, as long as their workload is not filesystem-intensive. Thin clients rely on the central server for the processing and thus require a fast server. When the central server is busy and slow, both kinds of clients will be affected, but thin clients will be slowed completely, whereas diskless nodes will only be slowed when accessing data on the server.
• Better multimedia performance. Diskless nodes have advantages over thin clients in multimedia-rich applications that would be bandwidth intensive if fully served. For example, diskless nodes are well suited for video gaming because the rendering is local, lowering the latency.
• Peripheral support Diskless nodes are typically ordinary personal computers or workstations with no hard drives supplied, which means the usual large variety of peripherals can be added. By contrast, thin clients are typically very small, sealed boxes with no possibility for internal expansion, and limited or non-existent possibility for external expansion. Even if e.g. a USB device can be physically attached to a thin client, the thin client software might not support peripherals beyond the basic input and output devices - for example, it may not be compatible with graphics tablets, digital cameras or scanners.

• The hardware is cheaper on thin clients, since processing requirements on the client are minimal, and 3D acceleration and elaborate audio support are not usually provided. Of course, a diskless node can also be purchased with a cheap CPU and minimal multimedia support, if suitable. Thus, cost savings may be smaller than they first appear for some organizations. However, many large organizations habitually buy hardware with a higher than necessary specification to meet the needs of particular applications and uses, or to ensure future proofing (see next point). There are also less "rational" reasons for overspecifying hardware which quite often come into play: departments wastefully using up budgets in order to retain their current budget levels for next year; and uncertainty about the future, or lack of technical knowledge, or lack of care and attention, when choosing PC specifications. Taking all these factors into account, thin clients may bring the most substantial savings, as only the servers are likely to be substantially "gold-plated" and/or "future-proofed" in the thin client model.
• Future proofing is not much of an issue for thin clients, which are likely to remain useful for the entirety of their replacement cycle - one to four years, or even longer - as the burden is on the servers. There are issues when it comes to diskless nodes, as the processing load is potentially much higher, thus meaning more consideration is required when purchasing. Thin client networks may require significantly more powerful servers in the future, whereas a diskless nodes network may in future need a server upgrade, a client upgrade, or both.
• Thin client networks have less network bandwidth consumption potentially, since much data is simply read by the server and processed there, and only transferred to the client in small pieces, as and when needed for display. Also, transferring graphical data to the display is usually more suited for efficient data compression and optimisation technologies (see e.g. NX technology) than transferring arbitrary programs, or user data. In many typical application scenarios, both total bandwidth consumption and "burst" consumption would be expected to be less for an efficient thin client, than for a diskless node.