You are browsing a read-only backup copy of Wikitech. The live site can be found at wikitech.wikimedia.org
Traffic cache hardware
Jump to navigation
Jump to search
This is an overview of our currently deployed and active cache hardware at the Traffic layer.
Hardware classes
We have purchased and retired multiple classes of server hardware over the years in staggered timeframes. In the general case we'll always have multiple overlapping classes of hardware as various warranty and support periods expire. These are the currently-active hardware configuration classes:
| Label | Model | CPU Type/Speed | Phys Cores | RAM | Cache storage | NIC speed, type, driver | DC Ops Config |
|---|---|---|---|---|---|---|---|
| F1 | Dell R440 | 2x Xeon Gold 5118 @ 2.3Ghz | 24 | 384 GB | 1x Samsung PM1725a 1.6TB NVMe (U.2 SFF) | 10G, BCM57412, bnxt | F-10G, +storage card |
| F2 | Dell R440 | 2x Xeon Gold 5118 @ 2.3Ghz | 24 | 384 GB | 1x Samsung PM1725b 1.6TB NVMe (HHHL Card) | 10G, BCM57412, bnxt | F-10G, +storage card |
| F3 | Dell R440 | 2x Xeon Gold 5118 @ 2.3Ghz | 24 | 384 GB | 1x Samsung PM1725b 1.6TB NVMe (HHHL Card) | 10/25G, BCM57412, bnxt | F-10G, +storage card, +10/25G NIC variant |
| F4-T | Dell R450 | 2x Xeon Gold 5318Y @ 2.1Ghz | 48 | 512 GB | 1x 6.4TB NVMe Card | 10/25G, BCM57414, bnxt | F (*) |
| F4-U | Dell R450 | 2x Xeon Gold 5318Y @ 2.1Ghz | 48 | 512 GB | 2x 6.4TB NVMe Card | 10/25G, BCM57414, bnxt | F (*) |
- F4: The new DC Ops Config F from mid-2022 is exclusive to these edge cache roles, and thus our F4 config using it includes the storage cards and NIC upgrades as part of its base definition.
Deployed hardware
Currently deployed hardware by data center and caching cluster.
| Data center | cache_text | cache_upload | total |
|---|---|---|---|
| eqiad | 8x F1 | 8x F1 | 16x F1 |
| codfw | 8x F2 | 8x F2 | 16x F2 |
| esams | 8x F2 | 8x F2 | 16x F2 |
| ulsfo | 8x F4-T | 8x F4-U | 16x F4 |
| eqsin | 8x F4-T | 8x F4-U | 16x F4 |
| drmrs | 8x F3 | 8x F3 | 16x F3 |
| total | 48x Fn | 48x Fn | 96x Fn |
Proposed FY22-23 changes + refreshes
Note as of 2022-12-08: the deployment of the new hardware in eqsin + ulsfo is now complete. Shipping some or all of the 8x F2/F3 nodes back to eqiad is still pending.
- ulsfo and eqsin get refreshed to new-standard 16xF4 config in first half of the FY.
- The 8x off-cycle (newer, still in warranty) F-nodes in ulsfo and eqsin are shipped to eqiad.
- Eqiad installs these into the new E+F rows for a number of reasons:
- Utilize the new rows in eqiad in general (more load/redundancy spread)
- Test impact of expanded server counts in general
- Re-use this good hardware instead of tossing it, so we don't waste it just for being off-cycle purchased
- Buys us time to push natural eqiad warranty refresh out another FY, spreading out refresh cycles better (too many this year!)
- Allow the F4 refreshes in ulsfo+eqsin to be whole-DC upgrades, since F4 enables whole-DC architecture changes in traffic routing.
- esams gets refreshed in Q4 to the same new F4 config as ulsfo+eqsin (we have some time and space to adjust this based on earlier outcomes if necc)
| Data center | cache_text | cache_upload | total | Note |
|---|---|---|---|---|
| eqiad | 8x F1 + 4x F2 | 8x F1 + 4x F3 | 16x F1 + 4x F2 + 4x F3 | Reinforced this FY |
| codfw | 8x F2 | 8x F2 | 16x F2 | no changes this FY |
| esams | 8x F4-T | 8x F4-U | 16x F4 | refreshed to F4 in Q4 |
| ulsfo | 8x F4-T | 8x F4-U | 16x F4 | refreshed to F4 in Q1 |
| eqsin | 8x F4-T | 8x F4-U | 16x F4 | refreshed to F4 in Q2 |
| drmrs | 8x F3 | 8x F3 | 16x F3 | no changes this FY |
| total | 52x Fn | 52x Fn | 104x Fn |