2018-11-18 06:52:42
I'm adding some benchmark and diagnostic features to Arktika
Fireduck
2018-11-18 07:31:00
and the benchmarking took down my vm server
Fireduck
2018-11-18 07:45:47
yeah, this ddr2 box can only do 18GB/s to ram
Fireduck
2018-11-18 07:45:48
heh
Fireduck
2018-11-18 07:50:11
sweet jesus, 142GB/s if I put in jvm ram vs tmpfs
Fireduck
2018-11-18 07:52:11
@Fireduck 18GB to 142GB? seriously?
Clueless
2018-11-18 07:52:26
yeah
Fireduck
2018-11-18 07:52:54
I wonder if there is something I should change with regards to filechannels
Fireduck
2018-11-18 07:52:59
it doesn't seem quite right
Fireduck
2018-11-18 07:53:25
the trouble it, tmpfs isn't a shit show like jvm heap
Fireduck
2018-11-18 07:53:40
so you can get a lot tigher, like 240gb in tmpfs on a 256gb system
Fireduck
2018-11-18 07:53:55
where with java heap, you are lucky to get maybe 180gb on a 256gb system
Fireduck
2018-11-18 07:54:17
that seems wrong
Clueless
2018-11-18 07:54:22
how the hell is there that much overhead
Clueless
2018-11-18 07:54:49
that is a separate question that i haven't figured out yet
Fireduck
2018-11-18 07:55:47
this is a big enough deal I need to test this in python.
Clueless
2018-11-18 09:48:59
it’s the GC, if it can avoid firing, it will
Rotonen
2018-11-18 11:31:05
Oh, that's a good point. Maybe you can set garbage collection to occur differently.
Clueless
2018-11-18 11:33:48
you can, that is a very wide permutation space of JVM tweaking
Rotonen
2018-11-18 16:24:42
also, it occurs to me that something might be optimizing and determined that since I didn't actually read the data after fetching it, why bother fetching it?
Fireduck
2018-11-18 16:24:53
so that 140GB/s number might be nonsense
Fireduck
2018-11-18 16:38:32
still 128GB/s after making it read things and do stuff
Fireduck