Report of the Athena/SGI Strategy Discovery Project

Contents

I. Summary

A. Project CEASE

The CEASE (Convening to Examine the Athena SGI Effort) team consisted of: Team mailing list: cease@mit.edu

Project notebook: http://web.mit.edu/cease/www/ (includes full charter and notes from meetings)

From late September through October, 1999, we met and gathered data to answer the discovery question:

Should we continue devoting the fiscal and human resources that are now going toward purchase and support for the SGI/Athena platform, or free them up for other uses?
within the confines of our scope statement:
This project is essentially a fact-finding mission, to gather information necessary for deciding on our future SGI/Athena strategy. In particular, this project will focus on:

Because this question impacts Athena development, cluster machine renewals, and departmental proposals, it is important to do this work now so that we are prepared to allocate staff resources efficiently, avoid disrupting the Athena release cycle, and generally be prepared for next year's machine purchases. We have frozen this year's departmental purchases for SGI O2s, and are accepting no new requests pending the outcome of this discovery effort.

B. Recommendations

Recommendation Principles:

  1. Phase out Athena-SGI support by summer 2003. Note that this coincides with SGI's announced end-date for active (as opposed to maintenance) support of the IRIX operating system and MIPS-based hardware, as well as the end of the normal 4-year life-cycle for machines placed in clusters summer 1999.
  2. For academic year 2000-2001, do NOT reduce the number of SGIs in clusters.
  3. Actively work with faculty relying on SGI-only software/features for courses to find non-SGI alternatives or develop other transition plans.
  4. Plan end-of-support date for private workstations to coincide with removal of machines from clusters.
  5. Make decisions on IRIX updates in the Athena release in line with usual release cycle, balancing development effort against importance of bug fixes or other gains.
  6. Keep Athena users informed of phaseout plans and transition paths (concentrating particularly on faculty and private workstation owners, but not to the exclusion of the general population).

Specific Recommendations:

C. Context (Summary)

Athena-SGI stakeholders can be grouped into three categories:
  1. Instructors and students using SGIs for course work
  2. IS staff involved in Athena-SGI support
  3. General Athena user base

SGI systems were originally deployed in the Athena Computing Environment as special purpose rather than general purpose client workstations. They were originally introduced to support courses requiring specific software applications available only on SGIs; over time, use of SGI-specific functionality has grown (both in courses and by the general user base), but software that requires SGIs continues to represent only a small fraction of Athena use in academics, isolated to a few specialized applications.

Supporting SGIs is expensive, both in terms of hardware and staff resources. An SGI O2 workstation costs 50% more than a comparably equipped (CPU speed, disk, memory) Sun Ultra5 and 15% more than the Sun Ultra10 with Creator 3D Graphics. We have found the SGI to be a more fragile platform in the clusters, with more down time and requiring more troubleshooting; as a vendor, SGI has been relatively uncooperative in working with our staff in tracking down bugs and problems when they arise.

The future of SGI appears to be uncertain and their strategy is unclear. We know that they plan to phase out the IRIX/MIPS low-end desktop platform and are concerned about continuing support from third-party software vendors. As we've seen in the past when DEC moved from supporting DEC Ultrix to Digital UNIX on Alphas, third-party vendors may stop issuing new versions for that platform, and may not guarantee that existing versions work across OS updates.

Meanwhile, IS is actively working to bring Linux and Windows NT into the Athena environment, and we need to think carefully about how many platforms we can support, given the constraints of staff resources and the limited number of seats in the clusters.

D. Scenarios (Summary)

CEASE examined various scenarios for the future of Athena-SGI, including discussion of development effort (IRIX updates), cluster renewals, and end-of-life for private workstations.

The most recently purchased cluster SGIs will reach the end of the usual 4-year life cycle in summer 2003, roughly coinciding with the time we have been led to believe that active support from SGI for the O2 platform will cease. This time frame could be used to gradually reduce SGIs from the environment as IRIX moves towards "maintenance mode", allowing us a natural transition period for dependent courses. In particular, it would be prudent to tie end-of-life for private workstation support to cluster removal dates, so that the commitment of staff resources does not extend beyond what is necessary for academic support.

The application of development resources to future IRIX updates should be decided in the usual Athena release forums and time frame, but some general considerations are discussed in part III-A.

II. Context

A. Athena-SGI Environment

Athena first started supporting SGIs in response to particular departments that use specific software applications only available on this platform (MSI and BioSym used by Chemistry, Chemical Engineering, and Materials Science, and Alias-Wavefront and Lightscape used by Architecture). We have since installed several SGI-only applications which do not appear to be used currently for courses. Appendix A summarizes usage data for Biosym/MSI, Alias-Wavefront, and other license-manager-logged applications (the Lightscape vendor dropped out of the SGI market a couple of years ago; it now runs only on NT).

CEASE identified the following other categories of course use:

Some faculty responded that they have moved away from previous SGI dependencies including Aero/Astro and Civil.

We are aware that there may be other courses using SGI-only homegrown or other non-commercial software installed in lockers which did not surface from our contacts, but believe the above list captures the vast majority of course dependencies, and that good publicity and general transition strategies for homegrown software will address any others. Appendix B outlines the process used to collect data on course use.

Independent of course use, the following benefits unique to SGIs were identified:

Athena-SGI deployment is as follows: 
			         year of purchase

			FY1997	FY1998	FY1999	FY2000	   Total   

  Public*		  44	  0	  22	  27	     93*
  Depts			  36	  3	  16	   5	     60
  Staff/SIPB		  12	  1	   9	   2	     24
  ----------------------------------------------------     ----
  Total			  92	  4	  47	  34	    177

  * Public SGIs: 3 in lecture halls and 90 in clusters (including 20 in 
    4-035, which doubles as a reservable classrooom).  There are no SGIs
    among Quickstations or in Dormitory clusters. 

For context, total Athena deploys are:

          Public Cluster      416   
          Eclassrooms          60 
          Lecture Halls        13  
          Quickstations        27 
          Dormitories          18 
          Depts	              395
	  Staff/SIPB          136
Notes on clusters:

B. Stakeholders

Stakeholder engagement:
  1. We compiled a list of courses, faculty, and departments known to use Athena-SGI, and sent a questionnaire to 30 faculty/TAs/administrators from this list (see Appendix B). In addition, Naomi Schmidt had already initiated discussions with several key faculty and has begun follow-up conversations with others on the basis of the questionnaire responses.

  2. The CEASE team included members from Athena Development and Academic Computing Support. We identified the following additional internal stakeholders:

    and a representative from each group was interviewed by a CEASE member to answer the following basic questions:

    Information from these interviews is incorporated throughout the report.

  3. Since the entire Athena user base is a stakeholder in any decision regarding Athena ports and clusters, part of the team's work was to consider how to assess and weigh use of SGIs outside of courses. Feedback from internal stakeholders provided valuable insight into several aspects of general Athena-SGI use detailed elsewhere in this report; it was agreed that it would be appropriate to mention this project at a recent public feedback session for Athena users (see below), but no more formal polling of the non-course user base was conducted.

    The Athena Feedback Forum was an open event held in late October to gather feedback from students on all aspects of Athena clusters. Attendance was low, but the comments on SGIs provided some anecdotal information:

C. Cost Analysis

Hardware

The unit price of an O2 as we have configured it has varied between $4690 and $5610 over FY1999 and FY2000. 
        FY1999  47 machines             $254,107
        FY2000  34 machines             $159,346

        Total (2 years, 81 machines)    $413,453
For comparison, below are prices for 81 machines from other vendors: 
        Sun Ultra 5     @ $2,867 ea     $232,227
        Sun Ultra 10    @ $3,733 ea     $302,373
        DELL Linux box  @ $2,000 ea     $162,000
There are no SGI-specific cluster infrastructure costs. (The same parts are used to anchorpad O2s as other workstations, but in a different configuration; the vendor did special design work on this, but did not charge us for it.)

SGI Software

SGI Varsity Packs: We paid $20,000 in FY2000 for additional licenses to match how many machines we have in the field (having underpaid previously). This is a one-time, per-machine right-to-use license; for private workstations, the cost is added to other Athena fees; for clusters and IS-support machines, the cost comes out of the ACS software budget. However, SGI allows us to "reuse" existing licenses from retired systems as we replace them with new SGIs, so we would incur additional costs only if we increase the number of SGIs.

Annual fees: We paid $10,000 in FY2000 to Unix-VMS-Help for our share of MIT's license entitling us to annual upgrades of Varsity Pack software. It's not clear what would happen if we had fewer SGI's in subsequent years. The per machine cost might go up. Also, if the machines were replaced by Suns, then our share of the Sun Software Library cost might go up.

Third-Party Software

Pricing for software on multiple platforms is generally based on number of users, not per platform, and thus replacing SGIs with an equivalent number of other machines would have a minimal effect on cost. As far as third-party software which runs only on SGIs (or which we currently only license for SGIs), whatever particular alternatives we adopt to support the courses depending on this functionality are not likely to cost less.

Support Costs Unique to SGIs

Supporting any additional Athena platform has numerous costs across IS groups which are not enumerated here; below are large-scale costs unique to SGI support.

Athena Development
Cluster Services
Athena Server Operations
Athena Consulting

D. Timeline for Decisions Impacting Academic Year 2000-2001


      Jan/Feb	Decide on IRIX version for Athena 8.4 
      		   (dev, release-team) 

      Mar/Apr	Cluster renewal discussions 
		   (vendors, Owls, ACMG, ITLT) 

      May	Decide on cluster renewals, place orders

E. SGI Plans

We have active discussions going on with SGI to enter into appropriate non-disclosure relations and to get the clearest and most accurate information of their plans. We will weigh the credibility of their plans based on past experience and demonstrated changes in SGI's behaviors. In the absence of explicit confirmation from SGI we have evolved the following interpretation based on public statments and our experiences:

We expect that over the next 12 to 18 months that we will be able to continue to buy a machine substantially equivalent to the 300 MHz R5200 O2s we have purchased thus far.

We expect that over that same time frame that additional speed and lower cost may be available in the O2 line.

However we also expect that soon SGI will offer as their preferred product in the low end, single user desktop workstation market a different product line than the O2, embodying a different CPU than the R5000 series MIPS processor and a different operating system than IRIX.

We expect, that if we do not opt to migrate to that new product line, the O2s would be actively supported with hardware spares, system software updates, and SGI-owned application software updates for at least four years after our purchase of those systems.

Even so, our overall expectation is that support for the O2 line would degrade over that time as SGI shifts its focus to its Linux/Intel strategy.

F. Third-Party Software Vendor Plans

We compiled a list of third-party software which we have installed only for SGI (aside from the SGI-supplied Varsity Pack applications). From this list, we eliminated those packages which were known to be available for Sun or Linux or for which we have source code. Summarized below are the responses we received from a survey sent to the vendors of the remaining packages. (See Appendix G for the survey questions.)

III. Scenarios

A. IRIX Support in the Athena Release

Athena 8.3 (the current release) uses IRIX 6.5.3m. It is highly unlikely that IRIX 6.6 would be released by SGI in time for inclusion in Athena 8.4 (to be deployed in summer 2000). Instead, the cost-benefit analysis likely to face Athena Development and the release-team will be whether to upgrade to a patch release such as 6.5.7m in order to address bugs, security issues, etc.

In another year, it should be clearer what practical impact SGI's phaseout plans have on IRIX releases between now and the time when they officially go into maintenance mode. In addition, third-party software vendor patterns should emerge, specifically whether new releases continue to appear for IRIX, and whether there are any dependencies on particular IRIX versions. Revisiting the state of course-dependencies at that time should help in the cost-benefit analysis for future IRIX updates.

At some point, Athena may elect to freeze IRIX support at a particular version and limit development effort strictly to maintenance.

B. Renewal Scenarios and Impact on Stakeholders

In FY2001, the cluster SGIs up for renewal (i.e., 4-year old machines) will be 44 Indys. The CEASE team explored five scenarios for these seats in academic year 2000-2001:
  1. Replace all Indys with non-SGIs.
  2. Keep half of the Indys, replace half with non-SGIs.
  3. Keep all of the Indys.
  4. Replace half with O2s, replace half with non-SGIs.
  5. Replace all Indys with O2s.
Appendix D provides more detail on all five scenarios.

Based on information from internal stakeholders and responses from faculty, we devised a grid to compare the relative impacts of the above scenarios on the major classes of stakeholders, and the budget (see Appendix E).

From this analysis, our recommendation falls between scenarios 2 and 3: do not buy any new O2s, but retain at least half of the Indys for a fifth year. If this recommendation is followed, then following the usual 4-year replacement cycle in subsequent years would gradually phase out all cluster SGIs as follows: 

	academic year   Indy   O2	total
	-------------  -----  ---	-----
	  1999-2000	  44   49	   93
	  2000-2001    22-44*  49	71-93*
	  2001-2002	   0   49	   49
	  2002-2003	   0   27	   27
	  2003-2004	   0    0	    0
* depends upon whether we choose to replace up to half of the FY1997 Indys with another platform (e.g., Linux or Pismere).

C. Private Workstation Support

Usual end-of-life for a particular platform is 18 months after machines of that type were removed from clusters, in order to allow a long grace period for private workstation owners to replace machines. Following this model would mean that Dec. 2004 would be the earliest date we could stop supporting Athena-SGIs. For example:
June 1999 Billing letter says we will desupport machines of type A in summer 2000.
Summer 1999 We remove all cluster machines of type A and offer to upgrade IS-granted machines based on proposals during 1999-2000.
June 2000 We bill for 6 months, saying things are only guaranteed to continue working until Dec. 31, 2000.

Alternatively, we could announce end-of-life for private machines to coincide with (or follow more closely after) removal from clusters; notices to departments could be issued earlier, to allow them ample time to plan for replacement platforms or reversion to vendor OS. This would allow us to provide the most robust support up to end-of-life, as private workstations would not lag appreciably behind the general field. There are a small number of machines (both ACS-granted and privately puchased) which fall into this category, i.e. which end their normal four-year life cycle at the same time that we plan to remove the remaining R5000 Indys (Summer 2001) or the remaining O2s (Summer 2003) from the public environment. ACS will confer with their owners to ensure a smooth transition (e.g. through early renewal, or reversion to vendor OS).

Another scenario we discussed was freezing development after cluster removals, but giving departments the option of staying with the old Athena release for an additional 6 months. This is not tenable as a support option, however, as SGIs are historically more sensitive than other platforms to infrastructure changes, and we could not guarantee that they would continue to interoperate in the Athena environment beyond the inclusion of IRIX in the development and testing process.

Appendix

Additional/future materials for Sponsor/ACMG:
Last modified: Fri Jan 14 13:33:59 2000
ajfox