Re: IDS 7.30 on Linux Tuning Benchmark?

From: Tim Schaefer <tschaefe@mindspring.com> > >I applied the "Kagel Krank" to the LRU's: LRUMAX 2 LRUMIN 1 For a load, I would rather set LRUMAX and MIN to high values, and only do a chunk write once all the data was in. This might require more buffers. But that's just me. :-) HTH. ______________________________________________________ Get Your Private, Free Email at http://www.hotmail.com

OK Clownster, I went from 2 and 1 to 98 and 1, did an onmode -c after each 10,000 rows. Nothing else changed in the onconfig, my buffers stayed the same, at 25% of Virtual Memory per the book. So... This gave me 2-second checkpoints, no FG's and all Chunk Writes:120135 No LRU writes. Very nice loading. I reduced the loading from 7:51 to 6:44, ( min:sec ) to now give me 1666 records/sec. This was interrupted though by the two-second checkpoints. I also tried taking out the onmode -c after each 10,000 row load, and doubled the buffers, but this caused long checkpoints averaging around 6 or 7 seconds. It did reduce the total load time further to 5:04. Very interesting, because the loads went like this: :06 :06 :14 < ckpt :06 :06 :05 :14 < ckpt Interesting, even at that it was a faster load. I guess this is the way to go for now. I've also loaded a 97,029 row loadfile in 17 seconds this way, with a 2-sec checkpoint. 5707 recs/sec. Very interesting. Also, if I load a series of these loadfiles, I get all chunk writes, no LRU writes, and checkpoints exactly even at 4 secs, with a 13 second average load time. ( 7463.76 recs/sec ) Hmmmmm... Linux ain't so bad I guess, but I really don't have a reference point other than XPS lately. It would be better to get a high-performance loader ported, so we could avoid the buffer problem altogether. This would eliminate having to chop up load files. Unless of course I'm missing something, I can't load large loads without incurring long transactions. BTW, Somebody implied lately that you went from vaudeville to the big top. Any truth to that? 8-) Obnoxio The Clown wrote: > > From: Tim Schaefer <tschaefe@mindspring.com> > > > >I applied the "Kagel Krank" to the LRU's: LRUMAX 2 LRUMIN 1 > > For a load, I would rather set LRUMAX and MIN to high values, and only do a > chunk write once all the data was in. This might require more buffers. But > that's just me. :-) > > HTH. > > ______________________________________________________ > Get Your Private, Free Email at http://www.hotmail.com -- - -- --- Tim Schaefer ---- tschaefe@mindspring.com --- http://www.inxutil.com -- -

Tim Schaefer wrote: [benchmark SNIPPED] > It would be better to get a high-performance loader ported, so we could > avoid the buffer problem altogether. This would eliminate having to chop up > load files. Unless of course I'm missing something, I can't load large loads > without incurring long transactions. How about dbload with the partial commit flag? Or my ul.ec binary file loader/unloader which also performs commits every N rows? Art S. Kagel

"Art S. Kagel" wrote: > > Tim Schaefer wrote: > [benchmark SNIPPED] > > It would be better to get a high-performance loader ported, so we could > > avoid the buffer problem altogether. This would eliminate having to chop up > > load files. Unless of course I'm missing something, I can't load large loads > > without incurring long transactions. > > How about dbload with the partial commit flag? Or my ul.ec binary file > loader/unloader which also performs commits every N rows? > > Art S. Kagel I'm interested in your EC program, will investigate shortly... Tim -- - -- --- Tim Schaefer ---- tschaefe@mindspring.com --- http://www.inxutil.com -- -