CIO Corner: The curious economics of centralized support



I CIO Corner: The curious economics of centralized support

II Dr. Jim Gray, Microsoft Research

III

IV Management introduction

In this CIO Corner Dr. Jim Gray offers insights into the costs of centralization. As he argues, system support costs (people) were once a relatively small element in the total cost of computing. Today that picture has altered radically – with profound implications, not least for distributed computing (including middleware) in general and network computers (NC) in particular.

The central arguments in this CIO Corner are taken from a paper written by Dr. Gray entitled Locally Served Network Computers. This was developed from discussions with Gordon Bell, Butler Lampson, and Phil Garrett. In Dr. Gray’s own words ‘Rick Rashid's skepticism about support costs challenged me to find quantitative arguments.’ This is the basis of his arguments here.

V The attraction of network computers

One of the main arguments in favor of network computers (NC) is that they will eliminate the care-and-feeding costs that dominate PC cost-of-ownership. Studies by Forrester Research and Gartner Group estimate that about 75% of a PC's cost-of-ownership is system administration - about $6K/year just to support a networked PC user. In addition, NCs might also reduce software costs (including middleware) – in that you buy for a central server and not for each PC.

NC advocates assure us that, since the NC is stateless, NCs will have no care-and-feeding. They argue all the support cost has moved to the centralized server – and then conveniently ignore those costs.

VI Flaws

There are at least two flaws in this pro-NC argument:

A bandwidth

B the costs of centralization.

The bandwidth problem is simple to explain. A stand-alone NC is useless. NCs are stateless, so browsers, applets, and data must all be downloaded from servers. A stand-alone NC is not a word processor, not a portable, and not a gameboy. To function, an NC must be connected to the network -- indeed it must be connected to a fast network if it is to match PC response times. Such wide-area network bandwidth is expensive today and will be expensive for years to come. To match PC response times over today’s public networks will require a hundred-fold reduction in communications prices to reach acceptable costs. This seems improbable to me.

The bandwidth issue is simple and easy to understand. The centralization issues are more complex. NCs postulate a central organization that provides the system management users now do themselves. Yes, this will be easier for users. But there is considerable evidence that this centralized management will be much more expensive and much less responsive:

C objective measures (see below) show that centralized servers require incredibly expensive staff (what are those IS guys doing???).

D experience shows that such staffs are often unresponsive to user needs (where was the applications backlog invented?).

NCs will fail if they require central servers managed by a party remote from the users needs. The popular NC vision postulates stateless NCs attached to giant servers run by a telco or Internet Service Provider -- the revenge of the mainframe. ‘You don’t want to manage your server? Why not let the phone company do it?’ I think there are two answers:

E convenience

F cost.

These two reasons explain why offices have PBXs, why Telco voice-mail has not caught on and why we avoid doing business with a computer center. If you are old enough, recall how horrible the data center was, recall how terrible timesharing was, recall how unresponsive the central IS organization was. I still hear about central IS organizations studying the benefits of Windows95. Meanwhile, most home users have made the conversion. I hope our children never experience these indignities. Convenience is the main motive for avoiding central IS organizations, but cost is easier to quantify, and motivate.

VII In perspective

You do not believe me? Consider the following. In the 1980s, data centers spent 60% on staff and 40% on equipment - very similar to the Forrester Research study of PC cost-of-ownership today ( research/cs/1995-ao/jan95csp.html). In the 1980's there were several rules-of-thumb to estimate staff sizes for a data center. My favorite rules were:

A a systems programmer per MIPS

B a data administrator per 10GB.

|Figure xxx. Federal Data Center IBM mainframe and compatible Operations Costs vs. Data Center Size. Large|

|data centers are more efficient (source Consolidation of Federal Data Centers, prepared by the Federal |

|Data Center Consolidation Committee, John R. Ortago Director, Council of Federal Data Center Directors) |

|Total MIPS |Average MIPS/machine |Yearly |Full Time |Estimated |

| | |$/MIPS |Staff / MIPS |staff/site |

|1-99 |36 |$127,500 |.92 |33 |

|100-199 |143 |$129,000 |.51 |72 |

|200-299 |258 |$100,000 |.76 |196 |

|300-399 |336 |$87,600 |.58 |195 |

|400 . . . |767 |$87,700 |.33 |253 |

In those days, an IBM MIPS cost $1M, 10GB of disk cost $1M and a staff person cost $60K/year. It was, therefore, common sense to hire people to manage a million dollars computing resource. The investment often made the equipment more productive and so paid the person’s salary.

Today those numbers have changed. It is now:

C About $8K per IBM CMOS MIPS (yes, PC MIPS are 1,000x cheaper)

D About $30K for 10GB of IBM disk (yes PC GBs are 10x cheaper)

E About $100K per staff person per year.

This illustrates what we all know – IBM hardware prices have declined by a factor of 300x since 1980; but labor costs have not changed much. It also highlights the incredible dis-incentive to buy mainframes.

It is hard to find quantitative data on the cost of managing a data center. Fortunately, the US government recently did a careful study of their data centers (Consolidation of Federal Data Centers, prepared by the Federal Data Center Consolidation Committee, John R. Ortago Director, Council of Federal Data Center Directors). This study examined 205 centers with a total staff of 12,900 people and a $1.2B budget. It examined centers of various sizes in detail and found the statistics shown in Figure Gr1

These numbers are astonishing. Two hundred people are employed to run a single 340 MIPS system. These are just the systems operators and administrators; they do not include either applications programmers or computer users. This is an huge ‘operating tax’.

Not unreasonably, the government concluded that big data centers get more MIPS per kilo-buck than little data centers. So, on the advice of the data center directors, the government mandated that each data center should have a minimum of 325 MIPS within two years. This will shrink the number of centers from 205 to about 50 (closing about 150 centers.) The remaining centers will average 200 operations staff each.

VIII Contrast the PC picture

My office has a 75 MIPS data center (a 25 MIPS laptop and a 50 MIPS tower). Our lab - of seven people - is a 600 MIPS data center.

Figure xxx suggests that our office needs a staff of 200 to mange it. Occasionally we wish for 200 people to manage these machines; but, alas we do not have budget or space for them. So, we do the work of 200 systems administrators in our spare time. Certainly it sometimes feels like we six are doing the work of 200, but I estimate we spend about 0.2 person years on it for our lab of six people. This is consistent with other studies showing that one support person can cover about 25 PC users.

This underlines one of the advantages of modern PC operating systems – they are a lot easier to install and administer then mainframe systems. 1,000x easier by this anecdotal evidence.

Now the mainframe people will argue that one IBM MIPS is worth more than one PC MIPS. They observe that an IBM MIPS is shared among many users. In contrast, a PC MIP is rarely used.

Let us be generous. Let us assume that the mainframe supports 10 users per MIPS. On this basis the operational staff (non-capital) costs of two data centers are:

A for the 340 MIPS mainframe data center, 200 staff at $100K/year each or $20M/year for 3,400 users. Or 6k$/user support cost.

• At 25 users per $100K/year support person, comes to 4k$/user support cost.

This means that user support costs are about the same for mainframe data centers as for PCs. Indeed, the data center calculations did not include the actual support and training costs for the end users, while the PC costs did.

IX The point

When someone tells you how cheap mainframe super-servers are going to be, you better hope they are made from GUI/NT clusters which are self-managing. Or, if they are using conventional mainframes, you had better prepare yourself for some massive bills from your friendly ‘NC Server Provider’.

If the NC is talking to the classic mainframe IS/IT structure, you can expect to be paying 6k$/user for the care and feeding of your server systems. Support costs will not be less expensive than PCs. Data center support costs will be comparable to the per-user PC support costs that NCs are supposed to eliminate. More likely they will be more costly than the PC model, for the reasons given above.

Centralized organizations have an qualitative downside -- they resist change. Their priorities are not your priorities. They decide to upgrade the system just when you need it most. They decide you do not really need that new version of DB2 or Oracle or 100 Mbps Ethernet or Windows95 or… After all, you got by with the old stuff yesterday. The galling thing is that they charge you a huge tax to make your life miserable.

To put it diplomatically, the main reason users welcomed PCs was because they perceived central organizations to be generally unresponsive to their (user) needs. Arguably, this was the main driver for first the mini-computer and then for PCs. Slow to respond, expensive service providers are not an obvious recipe for mass the success that some seem to expect of network computers.

X My model for the future

NCs are an evolution of the PC world, not an alternative. Today the PC, the printer, the portable, and the tower are each an independent computer. Each requires some setup and some management. It would be much simpler if the OS software automated this, making all these computers look like a single management entity. I should be able to just plug computers and applications into the network, much as I plug in telephones today.

The bandwidth requirements suggest that each home - or classroom or personal office - will need a local server. All the PCs, NCs and other computers in that area will be LAN connected to this server. You will be able to hang as many devices off this server as you like:

A in your office or classroom, the server will have an infra-red LAN to talk to all the devices there: your portables, the printers, the phones, ...

B in the home, the server will drive all the PCs/NCs spread throughout the house in place of telephones, radios, TVs, intercoms, and gameboys; the server will store your data and programs, give you a single ‘thing’ to manage and optimize your use of the expensive telco resource at the curb.

This is an evolutionary path where current PCs evolve into servers managing all the intelligent devices about to come into our lives. Some of those devices will be NCs. Of course this home, office, classroom, or group server will require management. The server software will manage everything that can be automated. But it will need guidance on policy issues like connections, connectivity, security, quotas, billing, etc. These new servers are likely to be built on plug-and-play PC operating system technology.

Some comparing the NC to X-terminals, and so conclude that stateless PCs will never catch on. I belive there will be stateless PCs in this new world. I believe X-terminals flopped because they were as expensive as a PC and they made you do business with some central authority (the IT problem). When there are several PCs in every room, the social and organizational barriers against NCs begin to evaporate. Few want a PC per family member per room. In such a world, the attraction of lower management costs quickly becomes a dominating argument for NCs.

In my world, the distinction between NCs and PCs is that NCs cannot operate detached from the network, while PCs may be used in stand-alone (mobile) devices. Otherwise, NCs and PCs will be equally easy to manage since the software will automate the management tasks we perform today.

XI Management conclusion

As Dr. Gray describes, stateless simple computers -- network computers -- are coming. But they will be LAN connected to local servers. By caching data and programs at the server, this architecture minimizes system management (including the middleware dimension) and optimizes use of expensive WAN bandwidth. The distinction between NCs and PCs is that PCs support stand-alone (mobile) operation.

From a middleware viewpoint this is to be welcomed for it will divide middleware into two categories:

A that used locally (for local connectivity of the various different devices and their applications)’ likely this will increasingly be a part of the local server operating system

• the external element, going from the local server to other local servers (be these at the dentist, insurance company, bank, your employer, friends and relatives, etc.).

This should become more manageable and at a price/performance level that is more acceptable than either of the existing extremes - the mainframe or the PC model.

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download