Compare commits
2 Commits
4736ee725d
...
25e5b28204
Author | SHA1 | Date |
---|---|---|
Vitaliy Filippov | 25e5b28204 | |
Vitaliy Filippov | 62fe6bf681 |
46
afr.js
46
afr.js
|
@ -42,57 +42,63 @@ function failure_rate_fullmesh(n, a, f)
|
|||
/******** PGS: EACH OSD ONLY COMMUNICATES WITH <pgs> OTHER OSDs ********/
|
||||
|
||||
// <n> hosts of <m> drives of <capacity> GB, each able to backfill at <speed> GB/s,
|
||||
// <k> replicas, <pgs> unique peer PGs per OSD
|
||||
// <k> replicas, <pgs> unique peer PGs per OSD (~50 for 100 PG-per-OSD in a big cluster)
|
||||
//
|
||||
// For each of n*m drives: P(drive fails in a year) * P(any of its peers fail in <l*365> next days).
|
||||
// More peers per OSD increase rebalance speed (more drives work together to resilver) if you
|
||||
// let them finish rebalance BEFORE replacing the failed drive.
|
||||
// let them finish rebalance BEFORE replacing the failed drive (degraded_replacement=false).
|
||||
// At the same time, more peers per OSD increase probability of any of them to fail!
|
||||
// osd_rm=true means that failed OSDs' data is rebalanced over all other hosts,
|
||||
// not over the same host as it's in Ceph by default (dead OSDs are marked 'out').
|
||||
//
|
||||
// Probability of all except one drives in a replica group to fail is (AFR^(k-1)).
|
||||
// So with <x> PGs it becomes ~ (x * (AFR*L/365)^(k-1)). Interesting but reasonable consequence
|
||||
// is that, with k=2, total failure rate doesn't depend on number of peers per OSD,
|
||||
// because it gets increased linearly by increased number of peers to fail
|
||||
// and decreased linearly by reduced rebalance time.
|
||||
function cluster_afr_pgs({ n_hosts, n_drives, afr_drive, capacity, speed, replicas, pgs = 1, degraded_replacement })
|
||||
function cluster_afr_pgs({ n_hosts, n_drives, afr_drive, capacity, speed, replicas, pgs = 1, osd_rm, degraded_replacement, down_out_interval = 600 })
|
||||
{
|
||||
pgs = Math.min(pgs, (n_hosts-1)*n_drives/(replicas-1));
|
||||
const l = capacity/(degraded_replacement ? 1 : pgs)/speed/86400/365;
|
||||
return 1 - (1 - afr_drive * (1-(1-(afr_drive*l)**(replicas-1))**pgs)) ** (n_hosts*n_drives);
|
||||
const resilver_disk = n_drives == 1 || osd_rm ? pgs : (n_drives-1);
|
||||
const disk_heal_time = (down_out_interval + capacity/(degraded_replacement ? 1 : resilver_disk)/speed)/86400/365;
|
||||
return 1 - (1 - afr_drive * (1-(1-(afr_drive*disk_heal_time)**(replicas-1))**pgs)) ** (n_hosts*n_drives);
|
||||
}
|
||||
|
||||
function cluster_afr_pgs_ec({ n_hosts, n_drives, afr_drive, capacity, speed, ec: [ ec_data, ec_parity ], pgs = 1, degraded_replacement })
|
||||
function cluster_afr_pgs_ec({ n_hosts, n_drives, afr_drive, capacity, speed, ec: [ ec_data, ec_parity ], pgs = 1, osd_rm, degraded_replacement, down_out_interval = 600 })
|
||||
{
|
||||
const ec_total = ec_data+ec_parity;
|
||||
pgs = Math.min(pgs, (n_hosts-1)*n_drives/(ec_total-1));
|
||||
const l = capacity/(degraded_replacement ? 1 : pgs)/speed/86400/365;
|
||||
return 1 - (1 - afr_drive * (1-(1-failure_rate_fullmesh(ec_total-1, afr_drive*l, ec_parity))**pgs)) ** (n_hosts*n_drives);
|
||||
const resilver_disk = n_drives == 1 || osd_rm ? pgs : (n_drives-1);
|
||||
const disk_heal_time = (down_out_interval + capacity/(degraded_replacement ? 1 : resilver_disk)/speed)/86400/365;
|
||||
return 1 - (1 - afr_drive * (1-(1-failure_rate_fullmesh(ec_total-1, afr_drive*disk_heal_time, ec_parity))**pgs)) ** (n_hosts*n_drives);
|
||||
}
|
||||
|
||||
// Same as above, but also take server failures into account
|
||||
function cluster_afr_pgs_hosts({ n_hosts, n_drives, afr_drive, afr_host, capacity, speed, replicas, pgs = 1, degraded_replacement })
|
||||
function cluster_afr_pgs_hosts({ n_hosts, n_drives, afr_drive, afr_host, capacity, speed, replicas, pgs = 1, osd_rm, degraded_replacement, down_out_interval = 600 })
|
||||
{
|
||||
let otherhosts = Math.min(pgs, (n_hosts-1)/(replicas-1));
|
||||
const otherhosts = Math.min(pgs, (n_hosts-1)/(replicas-1));
|
||||
pgs = Math.min(pgs, (n_hosts-1)*n_drives/(replicas-1));
|
||||
let pgh = Math.min(pgs*n_drives, (n_hosts-1)*n_drives/(replicas-1));
|
||||
const ld = capacity/(degraded_replacement ? 1 : pgs)/speed/86400/365;
|
||||
const lh = n_drives*capacity/pgs/speed/86400/365;
|
||||
const p1 = ((afr_drive+afr_host*pgs/otherhosts)*lh);
|
||||
const p2 = ((afr_drive+afr_host*pgs/otherhosts)*ld);
|
||||
const resilver_disk = n_drives == 1 || osd_rm ? pgs : (n_drives-1);
|
||||
const pgh = Math.min(pgs*n_drives, (n_hosts-1)*n_drives/(replicas-1));
|
||||
const disk_heal_time = (down_out_interval + capacity/(degraded_replacement ? 1 : resilver_disk)/speed)/86400/365;
|
||||
const host_heal_time = (down_out_interval + n_drives*capacity/pgs/speed)/86400/365;
|
||||
const p1 = ((afr_drive+afr_host*pgs/otherhosts)*host_heal_time);
|
||||
const p2 = ((afr_drive+afr_host*pgs/otherhosts)*disk_heal_time);
|
||||
return 1 - ((1 - afr_host * (1-(1-p1**(replicas-1))**pgh)) ** n_hosts) *
|
||||
((1 - afr_drive * (1-(1-p2**(replicas-1))**pgs)) ** (n_hosts*n_drives));
|
||||
}
|
||||
|
||||
function cluster_afr_pgs_ec_hosts({ n_hosts, n_drives, afr_drive, afr_host, capacity, speed, ec: [ ec_data, ec_parity ], pgs = 1, degraded_replacement })
|
||||
function cluster_afr_pgs_ec_hosts({ n_hosts, n_drives, afr_drive, afr_host, capacity, speed, ec: [ ec_data, ec_parity ], pgs = 1, osd_rm, degraded_replacement, down_out_interval = 600 })
|
||||
{
|
||||
const ec_total = ec_data+ec_parity;
|
||||
const otherhosts = Math.min(pgs, (n_hosts-1)/(ec_total-1));
|
||||
pgs = Math.min(pgs, (n_hosts-1)*n_drives/(ec_total-1));
|
||||
const resilver_disk = n_drives == 1 || osd_rm ? pgs : (n_drives-1);
|
||||
const pgh = Math.min(pgs*n_drives, (n_hosts-1)*n_drives/(ec_total-1));
|
||||
const ld = capacity/(degraded_replacement ? 1 : pgs)/speed/86400/365;
|
||||
const lh = n_drives*capacity/pgs/speed/86400/365;
|
||||
const p1 = ((afr_drive+afr_host*pgs/otherhosts)*lh);
|
||||
const p2 = ((afr_drive+afr_host*pgs/otherhosts)*ld);
|
||||
const disk_heal_time = (down_out_interval + capacity/(degraded_replacement ? 1 : resilver_disk)/speed)/86400/365;
|
||||
const host_heal_time = (down_out_interval + n_drives*capacity/pgs/speed)/86400/365;
|
||||
const p1 = ((afr_drive+afr_host*pgs/otherhosts)*host_heal_time);
|
||||
const p2 = ((afr_drive+afr_host*pgs/otherhosts)*disk_heal_time);
|
||||
return 1 - ((1 - afr_host * (1-(1-failure_rate_fullmesh(ec_total-1, p1, ec_parity))**pgh)) ** n_hosts) *
|
||||
((1 - afr_drive * (1-(1-failure_rate_fullmesh(ec_total-1, p2, ec_parity))**pgs)) ** (n_hosts*n_drives));
|
||||
}
|
||||
|
|
4
main.js
4
main.js
|
@ -31,7 +31,7 @@ class Calc extends preact.Component
|
|||
speed: st.speed/1000,
|
||||
ec: st.ec ? [ st.ec_data, st.ec_parity ] : null,
|
||||
replicas: st.replicas,
|
||||
pgs: 100,
|
||||
pgs: 50,
|
||||
degraded_replacement: st.eager,
|
||||
});
|
||||
this.setState(st);
|
||||
|
@ -67,7 +67,7 @@ class Calc extends preact.Component
|
|||
format4 = (n) =>
|
||||
{
|
||||
let p = Math.abs(n-(n|0)), m = 10000;
|
||||
while (p != 0 && p < 0.1)
|
||||
while (n < 1 && p != 0 && p < 0.1)
|
||||
{
|
||||
p = p*10;
|
||||
m = m*10;
|
||||
|
|
Loading…
Reference in New Issue