Speaker
Description
With the Run2025 for sPhenix, it comes the higher data throughput and data volume
requirements.
The sustained data throughput required for sPhenix2025 is 20GB/sec. Once started in
mid-April, this sustained data steam will be steadily constant with no breaks through December.
The projected data volume is 200PB.
In order to meet these data throughput and volume requirement, we must rebuild our
data storage archive systems…
1. Data movers.
a. Replace data movers with new PCIe-4 architecture and FC HBA’s. Increase the
number of data movers to nine servers from four.
2. Ethernet connecTons.
a. Replace NIC adapters with dual 100GbE PCIe-4.
3. Disk arrays.
a. With the NAND opTons too expensive, we decide to stay with HDD spindles
b. Increase disk arrays to nine units from three.
c. Increase each disk array mulTpath connecTons to four channels from two.
4. OS I/O related parameter tunings
5. Upgrade HPSS soYware to resolve occasional hanging processes on movers
6. Purchase addiTonal two 9-frame IBM tape libraries
a. To lower the costs, we decided to go for LTO9 technology, instead of enterprise
tape technologies.
b. Two 9-frame tape libraries are needed to meet the data volume requirements
c. To make the most usage of all tape drives, we deice to evenly strip injected data
across four tape libraries with 100 LTO9 drives (25 drive in each library).
7. Benchmark tesTng
a. We ran benchmark tesTng on exisTng 3 disk arrays with 4 movers and 56 tape
drives
b. With ongoing injecTon of 8.5GB/sec, the concurrent migraTon to tape is
11GB/sec.
c. With ongoing injecTon of 12GB/sec, the concurrent migraTon to tape is
9.9GB/sec.
d. The CPU usage on each mover is at 80%.
e. This benchmarking numbers give us confidence that our new configuraTons will
comfortably meet the 20GB/sec sustained throughput.
Conclusion:
With the new designs and fine-tunings, we have idenTfied a soluTon to the sPhenix2025
data requiremnets.
