专家问答：questions about linux clusters

[编者的话]

专家问答是Bio-IT的一个栏目，近期他们邀请了Linux NetworX 公司的技术官员Joshua Harr来回答有关于Linux Clusters的一些问题，这里节选了其中的一些问题和解答。

QUESTION:

I am interested in setting up a Linux cluster for our microarray core facilities. We can get a few old PCs. I wonder if any one has a good step-by-step guide for this purpose.

ANSWER :

Assembling spare PC equipment into a Beowulf cluster has a long successful history. Microarray analysis is a common application for these clusters. You can find benchmarking information for microarray analysis on a Beowulf cluster at: http://cmag.cit.nih.gov/Microarray.htm.

There are a number of good resources for building a cluster at: http://www.linuxhpc.org/pages.php?page=Documentation.

You can also find an active mailing list and wealth of other useful information at www.Beowulf.org. The mailing list is an excellent source of answers for beginners and experts alike.

Best of luck in your Beowulf adventure!

QUESTION:

Has anyone done a total-cost-of-ownership study on Linux clusters (vs. SGI, etc.)?

ANSWER :

To my knowledge, no one has performed a proper study of the total cost of ownership (TCO) of a Linux cluster versus alternatives. Some of the issues relating to total cost of ownership of a Linux cluster are covered in my response to the previous question. Some more specific work relating to reducing the TCO for a Linux cluster is covered in this article at the Cambridge Healthtech Institute: http://www.chiresource.com/newsarticles/issue12_2.asp

QUESTION:

We have a four-processor Sun and currently are serving approximately 40 users for high-throughput sequence analysis/annotation using GCG and in-house developed tools (mostly using BLAST and processing BLAST results).

In comparison to Unix (Sun Solaris), what are the hidden costs of operating a Linux cluster (apart from obvious issues about increased administration costs)? CPU-to-CPU, does the Linux solution require more floor space, cooling power, electricity, replacement parts, etc.?

If all you want to do is improve performance of BLAST, etc., is it more economical to invest in accelerator hardware such as Decypher?

ANSWER :

Overall, the total cost of ownership for Linux clusters relative to productive work performed is lower than any alternative I am aware of. The cost breakdown does differ from your current Sun system, however. Systems such as your Sun server tend to have very high initial up-front costs and high yearly maintenance costs with lower administration costs. Linux clusters generally have lower initial costs, lower priced yearly maintenance contracts, but higher administration costs (as you recognize in your question).

Overall, the largest cost comes from increased administration costs. There is a great amount of development work, by Linux NetworX and others, that aims to reduce this cost.

Linux clusters generally provide a significant reduction in the amount of required floor space compared to competitive systems. Several vendors offer systems that require 1U or less of rack space for the highest-performance servers (i.e., cluster nodes). The CPUs that are commonly used tend to draw a little more power and require more HVAC than your Sun system. The costs associated with power and cooling are negligible for most customers unless a very large system is deployed.

The largest factor in the reliability (i.e., maintenance costs) of Linux clusters is initial quality. The hardware used in Linux clusters changes rapidly and requires a dedicated testing effort to ensure that components work properly and work together in the system properly. A cluster from a vendor with good quality assurance procedures might be expected to have one to two RMA issues each year per fifteen nodes or so. The cost of replacement parts is generally covered by the vendor's support agreement, which is typically a fraction of the cost of a similar support agreement from a traditional high-end server vendor.

The greater administration costs and slightly greater maintainability are more than offset by the improvement in productive work that can be gained. It is very common to realize a 10-50x improvement in computational throughput per dollar spent compared to traditional high-performance computing equipment. A good vendor can help you benchmark your particular code on different platforms to help you determine the real-life improvement you will see. You can use this information to help you determine whether a Linux cluster is a good investment for you relative to other solutions, such as a BLAST accelerator. Also note that several groups are working on FPGA-based BLAST acceleration methodologies similar to DeCypher that can be deployed in a Linux cluster, taking advantage of the best of both solutions.

Answer Provided by: Joshua Harr

Joshua Harr is chief technology officer for Linux NetworX. Harr leads his company's effort to make clusters easier to manage, more powerful, and more reliable.