LSE (Linux Scalability Effort) Technical Meeting on 1/12/2001 ----------------------------------------------------------- Attendees: IBM: Tim Wright, Hans Tannenberger, Jonathan Lahr, someone for Hubertus from Watson SGI: John Wright, John Hawkes Intel: John B. Hansen, Tim Witham, Steve Carbonari Meeting Charter: This bi-weekly meeting is: - Technical Focus - Status Updates - Inject other agenda items by sending email to mailing list. Next Meeting is 1/26/01. Agenda: TBD Bridge numbers and Pass Code are same. Call 1-888-790-7156 Passcode:85875 Confirmation #: 8858784 Agenda / Minutes OSDL Overview by Tim Witham Web page is live with real data Building out the lab in next week and half. 1min overview: osdl is chartered with providing resources to developers working on enterprise capab. into Linux and OSS. Scalab, perf, clustering etc. eng. pool of large scale machines for people doing work, early prototyping, testing the work you do to make sure it works. Also for final integration and testing on large scale configurations. currently in rampup phase. Some initial systems coming in. 8ways builtup. Large 4ways with 4+gb memory. Hundreds of spindles eventually. current 50 2way machines as client loads. looking for feedback on types of systems you need for scalability. bought some early systems but buying more. Based on feedback of dev systems. 8by is biggest one they have. Big for current community. Web address: osdlab.org and osdl.net Jonathan Larhr doing work on 2way and 2.4 looks good. Tim: 2way issues looked at but > 2way may have problems. JW: many people working on scsci don't have the machines. Tim: many drives are fibrechannel for quick configuration. OSDL is engineering pool of systems for OSS developers. Rules (see web site): Must be oss license. Only OS supported is Linux Has to be a real project and code must exist where we can point people to it. Needs to be a web site or somewhere to see that code (hosted like SourceForge) to make sure code is really going public. Some exceptions but will go to board of directors. Looking at production qualified equipment vs. pre-production which is much harder (NDA and other board approval required) and its complicated in firewalling and giving certain users access etc. Lots of work is ok on straight prod. machineery. Equipment needs a defined programmatic interface. Generally, we don't sign nda or put hw in that isn't accessible to the general community on how to use it. JW: what about comparisons to non-oss stuff? e.g. closed sourc e driver which is already worked on oss in the community. Can you pull in the closed source module and compare it? Tim: we're not in the marketing/benchmarking business and rules for not doing publication and adds. Tim: the example seems reasonable but its touchy cuz the lab could get sued by anyone who's proprietary software. You the user must meet all the license restrictions. OSDL may have to kick you off to resolve something like that. JW: maybe the groundrule is only use the binary... Tim: you can use binaries for testing and load generation, you could use Oracle to generate a load in the kernel and then show before and after kernel loading data as long as you meet Oracle license requirements. Hans: What if you want to run benchmarks costing money? TPC-*.... Tim: How do you ....what costs you $? Running a TPC style or transaction style benchmark. TPC is fully audited. Hans: Dont you pay license fee just for running it? Tim: dont think so....sure we can get a basic setup and which db for that. You need permission from the DB vendor to publish anything. Its really complicated and can't say too much about. Tim Wright: use of open source db's...... Tim Witham: postgressql..... JW: What about Spec where you give them a license, even before auditing. Tim: Spec web ...but speccpu is more processor, not system. Should I get a specweb license? Very reasonable thing to do. JW: Some of the benchmarks hard to setup and run....do you offer services like a machine setup to run specweb aqnd throw your workload on it? Tim: not at this time, better approach is to find other people to help with that. Somebody donating time to help. Eventually, they will have skill sets to help on the more common things. current staff is to ensure smooth running systems and web based management and infrastructure. Don't have benchmark performance people. We have additional budget which could be steered that way. Whatever we decide. Tim: we're a resource, just the avail. of large scale equipment to non-coproate projects. We may have some regular projects on Sourceforge with people from the lab but those won't be OSDL projects. Jonathan: scheme for allocating resources? Tim: Its first come first serve. Will be a signup form - on the web. I need the following system config. for a certain amount of time. Jonathan: thats too clean....needs bureacracy. Tim: will need tuning for people who don't use their machines. People who break the rules etc. Its first come, first serve.... Linus or Alan may get some priority. If you need a 16way or 300 spindles for 3months, give Tim advance warning as much as possible. JW: When will ia64 be there...? Tim: needs it donated (pre-pro machine) or when he can purchase one w/out signing an NDA. Most people will have schedules on when that will happen. If other configs or instr. sets needed, let Tim know about that. Hans: machines pre-configured....what distros? Policy on distro's? Tim: ones we will have initially, Gang of 4 + Debian, RH, Turbo, Suse, Caldera, and Debian If you want to bring your own or send it...contact Tim. But, there are 1500 of them out there. Very neutral policy on that. JW: we can put our own stuff on it? Tim: yes, just a selection to add etc. but hope we don't need new drives etc. for Source. Will have mirrors of source stuff local also. Initial connection is a T1 and will add more bandwidth as we go. JW: how many real oss devs aware? Tim: Talked to Linus, folks at RH, guys at VA, all the major distros, ..... JW: any signups? Tim: one org is interested and everyone asking for access. Tim Wright: announce? Tim witham: big announce in next couple of weeks. Web site came up in Dec., next one in a week or two. Announce director and configs and stuff. Equipment is coming in and fair amount in already. We will be live with the equipment at the announce. Up and live by middle of next week. Tim Witham signing off: ------------------------------------ other topics --------------- John Wright: AIM benchmark....org was AIM technology and now defunct or bought by Network Assocf. When Ray Bryant and John Hawkes published at ALS.... IBM Lawyers talked to Network Assoc...and didn't care about the benchmark (cold shoulder). John Wright wants to pursue it as a useful benchmark. Tim Wright: send them a mail msg. that we will use it unless you say otherwise. JW: everyone has the source code but license and copyright is by AIM technology. TW: may be difficult to get permission from them. JW: can we find the people to get permission? Can someone from IBM.... AR-John Wright to send initial note to Ray and copy people on it. AR-Tim Wright can have someone look on IBM side. JW: send a note to Ray Bryant and followup that way. Hans: post it to lse-tech and see if anyone responds that way. ---------- other topic --------- scsi infrastructure stuff ------------ JW: mentioned IBM thinking about OSS their scsci infrast. stuff SGI getting theirs ported, don't know if will oss it.... community is talking about the issues..... JW: What does IBM have and what plans? TW: sequent scsci done long time ago but rich, good subsystem for err handling etc. have the original design docs (clean). They would need to work it but maybe not difficult to get that released...no-brainer to publish the docs and design specs. Could probably let you look at them. JW: on X-SCSI...enough difficulty....like XFS....to get it out there etc. Community doesn't want to accept something, wants to develop. TW: ....good for hacking on......... JW: current thinking....we will have X-scsi, but lower layers of driver are closed source due to encumbrances....intf. layer will probably be open source. doesn't get us to comm. acceptance Have some people willing to lead an effort and pull together the right pieces for a solid Linux subsystem. Can release pieces of Xscsi towards that and make it happen. Some people with lot of experiences and design docs but pushing xscsi is lot of work and may not get accepted. Leaning to....open what we can and work as a group to promote a new standard with best of both. TW: yes, the way to go...understand where you coming from. This way, might yield the best solution of all. JW: one thing can say...if half way down this path and XScsi is ok, JW can push much harder to get things opened up. TW: will look into and talk to Mike Anderson. JW: will try to stake out a section on the web site - a subproject which is tied very heavily into Scalability. TW: May need to be a separate project at some point but its sub-proj. of scalability definitely. Jw: community people want to make a library.....making legacy stuff inefficient.... ------- benchmark site ---------- JW: will put Jonathan's usage in there plus some other links. will clean up as well. Set of benchmarks for mysql and postgressql... Nice to get .....some stuff in a same framework and you can change and use with any db. TW: use ODBC? JW: not completely necessary...tpc-c setup that uses both Oracle and db2. JW: info on Wisconsin benchmark....as3ap.....coudn't find it.... discussed his recoding btrealworld benchmark at DG....started it again at SGI but dropped that.....would be nice to have something like that open source to use. TW: disk io generation....maybe port that to Linux....handy to have that around. JW: use DBENCH, with access pattern like netperf....(others)..... want to understand the differences of those. Need a comparison on those. Postmark? no, haven't used that. JW: also, not finding a lot of cache thrashing benchmarks..... TW: hogger.....same kind of idea.....some things in Pascal.... JW: before TPC-C at Sequent, synthetic benchmark workload with access patterns like TPC-C...and thrashed the cache...very synthetic.... TW: no more than TPC-C.... JW: we learned a lot 70% for something ez to setup. TW: really useful - could be JW: SGI looking at more scientific workloads also like MTI and OPENMT o NEXT MEETING ON 1/26/01 11-Noon PST