ScienceCloud hardware¶

Overview¶

ScienceCloud currently provides the compute instances and storage with the following grand totals:

Computing¶

nodes	virtual CPUs	total RAM
440	21'000	105 TB

GPU-Computing¶

Nodes are equipped with NVIDIA Tesla T4 and P4 GPUs.

nodes	GPUs	virtual CPUs	total RAM
22	44	608	2.5 TB

More info about GPUs on the ScienceCloud.

Storage¶

Block storage served by Cinder and object storage served by Swift:

type	raw capacity	usable capacity
Block storage	7.0 PB	2.3 PB
Object storage	1.7 PB	0.8 PB with replica-2 (or 1.2 PB with ec104)

Network¶

Every compute node has a non-blocking, redundant, 10gbps link to the internal network. This network is used to access the underlying storage infrastructure and provide network connectivity to the virtual machines.

The uplink to the University network is a redundant 20gbps link.

Flavors¶

Instances started on ScienceCloud land on different generations of CPUs depending on flavor. For available flavors, see the ScienceCloud flavor availability report.

Current ScienceCloud hypervisors¶

No. nodes	CPU type @ base clock	Base clock	Architecture	CPU/vCPU	RAM per node	Accelerators
96	2x Intel(R) Xeon(R) Gold 6126	2.60 GHz	Skylake	24/48	384 GB
2	2x Intel(R) Xeon(R) Gold 6126	2.60 GHz	Skylake	24/48	384 GB	2x nVidia P4
64	2x AMD EPYC 7702	2.0 GHz	Rome	64/64	512 GB
10	2x AMD EPYC 7702	2.0 GHz	Rome	64/64	512 GB	4 x nVidia T4

Both the instance root disk's data and any additionally attached volumes' data are stored on our CEPH cluster and are thus accessed via a redundant 20 or 50 gigabit network link.

To reduce fragmentation flavors are statically mapped to a specific hypervisor type. More specifically:

Flavor type	Hypervisor CPU
1cpu-4ram-hpcv3 2cpu-8ram-hpcv3 4cpu-16ram-hpcv3 8cpu-32ram-hpcv3 16cpu-64ram-hpcv3 32cpu-128ram-hpcv3	AMD EPYC 7702
8cpu-64ram-hpcv2-lmem 16cpu-128ram-hpcv2-lmem 32cpu-256ram-hpcv2-lmem	Xeon Gold 6126

There are a few cases in which it is desirable to have instances backed by the same generation of hardware (i.e. custom compiled software optimized with specific flags).

This can be accomplished on request by enabling specific flavors that are bound to a specific CPU generation. You should keep in mind that the binding solution will be more prone to issues during hardware maintenance as the instances cannot be moved easily from one host to another.

Flavor Availability Report¶

Direct link: https://cloud.s3it.uzh.ch/static/avail-public-flavors.html