1 2015-08-02 00:42:33 <denisx> I see that loading the chain is written to disk with fsync. would it be a problem to turn that off unitl the chain is completely loaded. it makes a huge difference performance wise
 2 2015-08-02 00:45:16 <midnightmagic> denisx: I bet they'd accept a patch for initial download for a configurable download fsync() based on amount of data.
 3 2015-08-02 00:46:30 <denisx> I did not touch any code, I just used zfs to turn it off while loading the chain
 4 2015-08-02 00:47:15 <midnightmagic> :-o
 5 2015-08-02 00:49:29 <denisx> and maybe it is not that easy because the write code is inside leveldb which knows nothing about the chain state
 6 2015-08-02 00:53:23 <Luke-Jr> denisx: pretty sure we delay fsync longer during IBD..
 7 2015-08-02 00:53:46 <Luke-Jr> eg, only every X blocks
 8 2015-08-02 00:55:45 <denisx> if I switch to sync=disabled the cpu rises from 130% to 190% for the bitcoin process
 9 2015-08-02 00:57:39 <denisx> ok, if I switch to sync=always the cpu goes down to 5%
10 2015-08-02 00:59:21 <denisx> is it correct that the IBD process is bound to two processes
11 2015-08-02 00:59:27 <denisx> it hits a wall at 200%
12 2015-08-02 01:54:12 <Luke-Jr> denisx: it tries to guess number of cores available
13 2015-08-02 01:54:39 <Luke-Jr> too many threads effectively kills performance on Windows :/
14 2015-08-02 01:54:59 <denisx> lock contention?
15 2015-08-02 02:06:43 <Luke-Jr> denisx: not performance of Bitcoin, but performance of the OS
16 2015-08-02 03:11:50 <Luke-Jr> hm, I wonder how many people out there think we devs *make* the rules, rather than merely discover and research them. O.o
17 2015-08-02 07:47:49 <phantomcircuit> <Luke-Jr> denisx: pretty sure we delay fsync longer during IBD..
18 2015-08-02 07:47:51 <phantomcircuit> yes
19 2015-08-02 07:49:54 <phantomcircuit> iirc sync=always means literally every write causes a sync