vitalif
/
vitastor
5
4
Fork 7
Code Issues 1 Pull Requests 2 Actions Releases 53 Wiki Activity

One more experiment with cluster AFR%

Vitaliy Filippov 2020-09-27 19:42:42 +03:00
parent f460d8c1c8
commit 8e36f04482
1 changed files with 33 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
mon/afr.js
View File

@ -4,6 +4,7 @@
// License: VNPL-1.0 (see README.md for details)
module.exports = {
cluster_afr_domains,
cluster_afr,
failure_rate,
c_n_k,
@ -14,6 +15,38 @@ console.log(100*cluster_afr(1000, 0.03, 1, 3), '%');
console.log(100*cluster_afr(5, 0.1, 1, 2), '%');
console.log(100*cluster_afr(14, 0.01, 1, 2), '%');
console.log(100*cluster_afr(100, 0.03, 1, 2), '%');
console.log(100*cluster_afr_domains(10, 10, 0.03, 0.08, 8000, 0.02, 10/8, 2, 70), '%');
console.log(100*cluster_afr_domains(10, 100, 0.03, 0.08, 8000, 0.02, 10/8, 2, 70), '%');
// Estimate AFR of the cluster, taking failure domains and server failures into account
// n_nodes - number of failure domains
// drives_per_node - number of drives in a failure domain
// afr_drive - AFR of a single drive
// afr_node - AFR of a single failure domain (server) not counting its drives' AFR
// osd_size - drive size in GB
// rebalance_speed_per_drive - rebalance speed per drive in GB/s
// network_speed - network speed in GB/s
// k - replication factor
// pg_per_osd - average PG per osd
function cluster_afr_domains(n_nodes, drives_per_node, afr_drive, afr_node, osd_size, rebalance_speed_per_drive, network_speed, k, pg_per_osd = 100)
{
const any_node_drive_afr = failure_rate(drives_per_node, afr_drive, 1);
// Logic behind this: one failed drive is expected to backfill to (pg_per_osd / 2) peers if we expect 50% collisions
// Which in fact is more correct for Ceph than for Vitastor because Vitastor's PG distribution isn't usually random!
const drive_rebalance_speed = Math.min(rebalance_speed_per_drive * Math.min(drives_per_node, pg_per_osd / 2 / n_nodes), network_speed);
const drive_rebalance_days = osd_size / (drive_rebalance_speed*n_nodes) / 86400;
const node_rebalance_speed = Math.min(rebalance_speed_per_drive * Math.min(drives_per_node, pg_per_osd / 2 * drives_per_node / n_nodes), network_speed);
const node_rebalance_days = osd_size*drives_per_node / (node_rebalance_speed*(n_nodes-1)) / 86400;
let p = 0;
for (let i = 0; i < n_nodes-(k-1); i++)
{
p += any_node_drive_afr * (1-any_node_drive_afr-afr_node)**i
* failure_rate(n_nodes-i-1, (any_node_drive_afr+afr_node)*drive_rebalance_days/365, k-1);
p += afr_node * (1-any_node_drive_afr-afr_node)**i
* failure_rate(n_nodes-i-1, (any_node_drive_afr+afr_node)*node_rebalance_days/365, k-1);
}
return p;
}
// Estimate AFR of the cluster (not taking failure domains into account)
// n - number of drives