Hardware Requirements¶
The purpose of this section is to describe processor, storage, and network requirements to deploy a DAOS system.
Deployment Options¶
A DAOS storage system is deployed as a Pooled Storage Model. The DAOS servers can run on dedicated storage nodes in separate racks. This is a traditional pool model where storage is uniformly accessed by all compute nodes. In order to minimize the number of I/O racks and to optimize floor space, this approach usually requires high-density storage servers.
Processor Requirements¶
DAOS requires a 64-bit processor architecture and is primarily developed on Intel x86_64 architecture. The DAOS software and the libraries it depends on (e.g., ISA-L, SPDK, PMDK, and DPDK can take advantage of Intel Streaming SIMD (SSE) and Intel Advanced Vector Extensions (AVX).
Some success was also reported by the community on running the DAOS client on 64-bit ARM processors configured in Little Endian mode. That being said, ARM testing is not part of the current DAOS CI pipeline and is thus not validated on a regular basis.
Network Requirements¶
An RDMA-capable fabric is preferred for best performance. The DAOS data plane relies on OFI libfabric and supports OFI providers for Ethernet/tcp and InfiniBand/verbs. Starting with a Technology Preview in DAOS 2.2, UCX is also supported as an alternative network stack for DAOS. Refer to UCX Fabric Support (DAOS 2.2 Technology Preview) for details on setting up DAOS with UCX support.
DAOS supports multiple network interfaces on the servers by binding different instances of the DAOS engine to individual network cards. DAOS can support multiple network interfaces on the clients, by assigning different client processes on the node to different network interfaces. Note that DAOS does not support network-level striping over multiple network interfaces, so a single client process will always use a single network link.
The DAOS control plane provides methods for administering and managing the DAOS servers using a secure socket layer interface. Management traffic between clients and servers uses IP over Fabric. On large clusters, however, administration of DAOS servers typically uses an additional out-of-band network connecting the nodes in the DAOS service cluster.
Storage Requirements¶
DAOS requires each storage node to have direct access to storage-class memory (SCM). While DAOS is primarily tested and tuned for Intel Optane^TM^ Persistent Memory, the DAOS software stack is built over the Persistent Memory Development Kit (PMDK) and the Direct Access (DAX) feature of the Linux operating systems as described in the SNIA NVM Programming Model. As a result, the open-source DAOS software stack should be able to run transparently over any storage-class memory supported by the PMDK.
The storage node can optionally be equipped with NVMe (non-volatile memory express)[^10] SSDs to provide capacity. HDDs, as well as SATA and SAS SSDs, are not supported by DAOS. Both NVMe 3D-NAND and Optane SSDs are supported. Optane SSDs are preferred for DAOS installation that targets a very high IOPS rate. NVMe-oF devices are also supported by the userspace storage stack but have never been tested.
A minimum 6% ratio of SCM to SSD capacity will guarantee that DAOS has enough space in SCM to store its internal metadata (e.g., pool metadata, SSD block allocation tracking). Lower ratios are possible, but the amount of SCM required will depend on the usage patterns of the applications accessing the DAOS storage. Since DAOS uses the SCM for its metadata, if the ratio is too low, it is possible to have bulk storage available but insufficient SCM for DAOS metadata.
For testing purposes, SCM can be emulated with DRAM by mounting a tmpfs filesystem, and NVMe SSDs can be also emulated with DRAM or a loopback file.
Storage Server Design¶
The hardware design of a DAOS storage server balances the network bandwidth of the fabric with the aggregate storage bandwidth of the NVMe storage devices. This relationship sets the number of NVMe drives depending on the read/write balance of the application workload. Since NVMe SSDs have read faster than they write, a 200Gbps PCIe4 x4 NIC can be balanced for read only workloads by 4 NVMe4 x4 SSDs, but for write workloads by 8 NVMe4 x4 SSDs. The capacity of the SSDs will determine the minimum capacity of the Optane PMem DIMMs needed to provide the 6% ratio for DAOS metadata.
CPU Affinity¶
Recent Intel Xeon data center platforms use two processor CPUs connected together with the Ultra Path Interconnect (UPI). PCIe lanes in these servers have a natural affinity to one CPU. Although globally accessible from any of the system cores, NVMe SSDs and network interface cards connected through the PCIe bus may provide different performance characteristics (e.g., higher latency, lower bandwidth) to each CPU. Accessing non-local PCIe devices may involve traffic over the UPI link that might become a point of congestion. Similarly, persistent memory is non-uniformly accessible (NUMA), and CPU affinity must be respected for maximal performance.
Therefore, when running in a multi-socket and multi-rail environment, the DAOS service must be able to detect the CPU to PCIe device and persistent memory affinity and minimize, as much as possible, non-local access. This can be achieved by spawning one instance of the I/O Engine per CPU, then accessing only the persistent memory and PCI devices local to that CPU from that server instance. The DAOS control plane is responsible for detecting the storage and network affinity and starting the I/O Engines accordingly.
Fault Domains¶
DAOS relies on single-ported storage massively distributed across different storage nodes. Each storage node is thus a single point of failure. DAOS achieves fault tolerance by providing data redundancy across storage nodes in different fault domains.
DAOS assumes that fault domains are hierarchical and do not overlap. For instance, the first level of a fault domain could be the racks and the second one, the storage nodes.
For efficient placement and optimal data resilience, more fault domains are better. As a result, it is preferable to distribute storage nodes across as many racks as possible.