7.23 loses pages off the buffer hash chain?

Informix On-Line Dynamic Server 7.23.UC9 HP K420 4-way, HP-UX 10.10, 2 GB RAM, one dbspace disk. We're trying to keep a 780 MB database fully cached, and have allocated 550000 buffers (1100 MB) for it. We're also trying to keep it clean so checkpoints are short, so we've got 3 full-time page cleaner threads running. LRUS is set to 16 (though we've tried 17 to avoid another possible bug in the hashing). Only one CPU VP is configured to run. We pre-load the cache, then start the application running. Application does a lot of singleton selects, then a tablespace BLOB update and counter update (this is the hot row) on each user invocation. onstat -h (apart from the bug in calculating the number of hashedbuffers) shows a steady loss of pages from the hash chains (I calculate this by taking the number of 2-page chains, multiplying by 2 and adding the number of 1-page chains to figure out how many are actually hashed). Cached reads falls gradually, and cached writes and bufwaits start increasing (but we know we're I/O bound on the one disk; we're working that as a separate issue). INFORMIX-OnLine Version 7.23.UC9 -- On-Line -- Up 01:40:41 -- 1257824 Kbytes buffer hash chain length histogram # of chains of len 800981 0 125505 1 122090 2 1048576 total chains 247595 hashed buffs 550000 total buffs calculated 369685 INFORMIX-OnLine Version 7.23.UC9 -- On-Line -- Up 01:41:11 -- 1257824 Kbytes buffer hash chain length histogram # of chains of len 800982 0 125562 1 122032 2 1048576 total chains 247594 hashed buffs 550000 total buffs calculated 369626 ( lost 59 ) INFORMIX-OnLine Version 7.23.UC9 -- On-Line -- Up 01:41:43 -- 1257824 Kbytes buffer hash chain length histogram # of chains of len 800976 0 125648 1 121952 2 1048576 total chains 247600 hashed buffs 550000 total buffs calculated 369552 ( lost 74 ) The profile: INFORMIX-OnLine Version 7.23.UC9 -- On-Line -- Up 01:41:11 -- 1257824 Kbytes Profile dskreads pagreads bufreads %cached dskwrits pagwrits bufwrits %cached 71848 71848 10724659 99.33 481812 560171 503075 4.23 isamtot open start read write rewrite delete commit rollbk 17146594 1013977 1413979 1256115 45 197190 0 98622 0 ovlock ovuserthread ovbuff usercpu syscpu numckpts flushes 0 0 0 1972.43 461.92 162 324 bufwaits lokwaits lockreqs deadlks dltouts ckpwaits compress seqscans 75276 46427 7711793 0 0 162 20608 14 ixda-RA idx-RA da-RA RA-pgsused lchwaits 26805 0 0 26811 43606 The suspected dropping of pages from the hash gets better with one cleaner, but then our checkpoint times get too long. Has anyone seen this before? Can it be explained? (The case with Informix is pending ...) Can it be configured around? Regards, Jim -- W. Jim Jordan, Nortel Networks, Stop 313 Qualicum, PO Box 3511 Stn C, Ottawa, ON K1Y 4H7 Canada wjjordan@nortelnetworks.com I do not speak for Nortel Networks. I do not do business with spammers.