1 |
On Sunday, 12 July 2020 12:38:22 BST William Kenworthy wrote: |
2 |
> On 12/7/20 6:03 pm, Nikos Chantziaras wrote: |
3 |
> > echo bfq > /sys/block/sda/queue/schedule |
4 |
> |
5 |
> Thanks for the hints, |
6 |
> |
7 |
> ive gone with schedtool and ionice for now (seems to be working) and |
8 |
> will configure that as the defaults when this run finishes. I have not |
9 |
> built the bfq scheduler in this kernel so will give that a try later - |
10 |
> its currently using mq-deadline with a WD blue SSD (will bfq be better |
11 |
> than a deadline on an ssd? - will try it and see). It has 4g ram and 4g |
12 |
> swap with about half in use - in normal running swap isn't used much, if |
13 |
> at all. |
14 |
|
15 |
BFQ manages the I/O pipe more effectively/efficiently, so if the pipe gets |
16 |
full due to e.g. heavy swapping, you'll see an improvement. However, with an |
17 |
SSD the improvement will be less noticeable than with a spinning disk and with |
18 |
an NVMe even less. |
19 |
|
20 |
|
21 |
> As well as emerge, the io load is also coming from the network as I have |
22 |
> 3x1g bonded interfaces routing busy vlans including a moosefs SAN and a |
23 |
> 100mhz uplink which was suffering delays and timeouts. The root cause |
24 |
> is trying to do too much with too little - :) |
25 |
> |
26 |
> BillK |
27 |
|
28 |
Network storage will be managed by the remote kernel, so your atom's scheduler |
29 |
won't help with that. |