G
George Potemkin
Guest
The real reason to start this thread is the performance issue that the same customer has in their main application. The bottleneck is MTX latch. Db is connected by approximate 4 thousand users that are creating 200-500 active transactions opened at the same time. 100+ commits per sec. Rollback of just one large transaction or the massive record deletion done by a couple sessions at the same time cause the end-users to complain about performance. MTX latch becomes busy 60% of time. Bigrow tests show 4 MB/sec of the unbuffered synchronous writes. I would expect 2-3 times higher rate. Unfortunately I don't know when the customer run the tests and how busy were the disks at that moment. There are 2 sec pauses during checkpoints. Though it's maybe normal for -B 2,000,000 + -B2 1,000,000. No. Time Len Freq Dirty CPT Q Scan APW Q Flushes Duration Sync Time 406 12:27:20 248 259 27338 21387 7490 553 0 1.75 0.64 407 12:31:39 200 216 26326 22649 4156 475 0 1.82 0.69 408 12:35:15 222 237 23782 18915 5772 384 0 2.52 1.40 409 12:39:12 223 235 23843 20442 3085 462 0 1.75 0.64 410 12:43:07 226 236 25166 22218 2760 183 0 1.86 0.75 411 12:47:03 208 217 26666 24048 2721 414 0 1.99 0.88 412 12:50:40 187 196 24945 23267 3109 147 0 2.10 0.98 413 12:53:56 227 237 22586 19827 5270 657 0 2.11 0.00 414 12:57:53 161 0 27228 18055 4600 239 0 2.07 0.95 My main question: can the slow disks cause the contention for MTX latch?
Continue reading...
Continue reading...