1
0
Files
2022-09-29 17:59:04 +03:00

148 lines
5.1 KiB
Plaintext

libirixpmda
X[Y] means assigned to X, status Y (blank=not started, D=development,
T=testing, X=done)
Lego documents, and info sources
. I have paper copies of the Hub, Bridge and Xbow specs
. others from http://b7.asd/doc/ (access controls are in effect here,
but they have been removed for gonzo, bozo, sandpit and moomba)
. general ficus stuff, benf@engr
. hardware graph, len@engr.
. See /hosts/sandpit/usr/people/markgw/Doc/lego/len/naming.ps
. xbow biyabani@engr
. See http://holodec.engr.sgi.com/~biyabani/README.xbmon
-------
M-05[T] base disk stats, jeffreyh@engr
. Needs testing on larger machines, multiple IO4s and/or SCIPs disks.
. Ania to do testing
?-08[ ] multiple IO4s and multiple SCIPs, tomf@trevose, cwr/jeremy
A-10[D] QA scripts for Irix-QA-Release people to use
. automated scripts that sanity check libirixpmda instrumentation
against "reasonable" values and other irix tools.
# Deferred beyond ficus FCS (i.e. irix6.4-ssg)
DM-20[D] page migration, numa/nstats(1), lfsh@engr
. need clarification of some some metrics ... some seem bogus
or not implemented
. per-node memory usage metrics
%free + %local + %repl + %lent == 100% ???
See sysmp(MPSA_NODE_INFO) in sysmp.h
- need to understand _what_ these mean
- numa_stats.pagealloc_lent_memory is in pages and is broken;
nar@engr posted bug #434711
M-25[D] hub (I/O counters broken in current version of hardware)
. started trying to figure this out
need feedback on how to access hub registers in user mode.
. MD (Memory Directory) registers ... 4-bit perfsel and 1-bit
perfenable ... 6 x 20-bit registers in silicon ... peg at
max
. /dev/mem reader/writer -- NO
. Jim Laudon is interested
. no h/w interrupts ... kernel API under development [e-mail 10/14]
?-12[ ] needs vehicle to explain how to update PMNS for new IRIX metrics
using /var/pcp/pmns/Rebuild
(a) at the PCP collector site, and
(b) at the PCP monitor site
. eoe relnotes? pcp_origin relnotes?
?-09[ ] If this happens, drop o32 format libirixpmda, else drop n32 format
From: Ben Fathi
Date: Wed Sep 4
Subject: Re: 32-bit kernels from ficus tree
yes, IP22. currently o32, supposed to be n32 by release time.
Also mail from olson@engr: more than IP22 ... but same o32/n32 story
. need to track this, and if n32 special care for QA'ing libirixpmda
. are both o32/n32 installed on a 32-bit kernel (Speedo?)
?-??[X?] irix.mem units, bug #365244
. issue is mem op counters versus mem space measures ... need
to check kernel source and review help text,
e.g. irix.mem.system.sptfree
. this has been fixed in 6.2 according to closure text for #365244.
MD-??[D] PCP protocol extensions
. protocol version (and byte order?)
. authenication for unlicensed pmcd use by trusted irix apps
e.g. 3Dconsole, netstat, dkstat, etc. (leedom@asd.sgi.com)
. Test implementation, as specified in mark's mail 30Sept96.
?-??[ ] environ pmda
. needs review w.r.t. changes to syscntrld in ficus
. lego systems may have multiple fans in multiple cabinets.
. wait for syscntrld
?-??[ ] network stats, netstat(1), fischer@engr, vjs@engr
. convert to avoid /dev/kmem for network interfaces
. rfe #430097 to add multicast per-interface byte snd/rcv.
(currently needs to walk kmem if we do this).
?-??[ ] re-instate gfx instr (see gr_osview, Jonathan's mail, reuel@asd)
also be aware of dmedia stuff in bonsai
. osview and gr_osview still use kmem readers.
?-??[ ] xlv instr: according to geoff@engr this will be for XLV^2 which
will not happen before kudzu MR
?-??[ ] SIGABRT -> named(1)
?-??[ ] HiPPI, by pass bug, fddistats(1)
?-??[ ] smt/apps/smtstat(1)
?-??[ ] cachefs, sun/cachefs/tools/cfsstat(1)
?-??[ ] grio, xfs/grio/griostat.h
?-??[ ] array, array/astat, astat(1)
. syssgi(SGI_ENUMASHS) call to enumerate all ashes, rdb@homegrown.
?-??[ ] tlbstats
?-??[ ] token ring, tokenring/trvis
?-??[ ] RSVP, bsd/rsvp/tools/rstat.c
?-??[ ] disk stats
. node-level aggegation? <= disk:node map is computable
?-??[ ] reconcile, standardize and clean up startup messages when
libirixpmda has a problem, e.g. on an IP19
?-??[ ] irix.kernel.all.cpu.active
. irix.kernel.all.cpu.util would be better (intended to be the sum
of all the !idle && !wait time
. scale to max of 100%, independent of no. of CPUs
. also do percpu version
?-??[ ] memory tiles (Moosehead only?)
?-??[ ] xlv failover, SP/A and SP/B stats before and after failover
. "pmstore irix.pmda.reset 1" in xlv failover?
?-??[ ] bridge
. no h/w instr => no kernel ctrs? [e-mail 10/15]
?-??[ ] fiber channel disks, raster@engr
. between Lego FCS and SCS
. what is the IRIX delivery vehicle? ... IRIX 6.4 patch
?-??[ ] xbow, xbstat, biyabani@engr
. conflict with rtmon use (event-trace) => T5 evctr-like scheme for
process-specific vs system-level usage being mutually exclusive
. email dialog established with biyabani@engr, see
http://holodec.engr.sgi.com/~biyabani/README.xbmon
?-??[ ] disabled CPUs
. is map xlate correct hinv-only IP19 and HWG systems
. does it work for CPU disabled in PROM?