.. _hardware-recommendations: ========================== Hardware Recommendations ========================== Ceph was designed to run on commodity hardware, which makes building and maintaining petabyte-scale data clusters economically feasible. When planning out your cluster hardware, you will need to balance a number of considerations, including failure domains and potential performance issues. Hardware planning should include distributing Ceph daemons and other processes that use Ceph across many hosts. Generally, we recommend running Ceph daemons of a specific type on a host configured for that type of daemon. We recommend using other hosts for processes that utilize your data cluster (e.g., OpenStack, CloudStack, etc). .. tip:: Check out the `Ceph blog`_ too. CPU === CephFS metadata servers are CPU intensive, so they should have significant processing power (e.g., quad core or better CPUs) and benefit from higher clock rate (frequency in GHz). Ceph OSDs run the :term:`RADOS` service, calculate data placement with :term:`CRUSH`, replicate data, and maintain their own copy of the cluster map. Therefore, OSD nodes should have a reasonable amount of processing power. Requirements vary by use-case; a starting point might be one core per OSD for light / archival usage, and two cores per OSD for heavy workloads such as RBD volumes attached to VMs. Monitor / manager nodes do not have heavy CPU demands so a modest processor can be chosen for them. Also consider whether the host machine will run CPU-intensive processes in addition to Ceph daemons. For example, if your hosts will run computing VMs (e.g., OpenStack Nova), you will need to ensure that these other processes leave sufficient processing power for Ceph daemons. We recommend running additional CPU-intensive processes on separate hosts to avoid resource contention. RAM === Generally, more RAM is better. Monitor / manager nodes for a modest cluster might do fine with 64GB; for a larger cluster with hundreds of OSDs 128GB is a reasonable target. There is a memory target for BlueStore OSDs that defaults to 4GB. Factor in a prudent margin for the operating system and administrative tasks (like monitoring and metrics) as well as increased consumption during recovery: provisioning ~8GB per BlueStore OSD is advised. Monitors and managers (ceph-mon and ceph-mgr) --------------------------------------------- Monitor and manager daemon memory usage generally scales with the size of the cluster. Note that at boot-time and during topology changes and recovery these daemons will need more RAM than they do during steady-state operation, so plan for peak usage. For very small clusters, 32 GB suffices. For clusters of up to, say, 300 OSDs go with 64GB. For clusters built with (or which will grow to) even more OSDS you should provision 129GB. You may also want to consider tuning settings like ``mon_osd_cache_size`` or ``rocksdb_cache_size`` after careful research. Metadata servers (ceph-mds) --------------------------- The metadata daemon memory utilization depends on how much memory its cache is configured to consume. We recommend 1 GB as a minimum for most systems. See ``mds_cache_memory``. OSDs (ceph-osd) --------------- Memory ====== Bluestore uses its own memory to cache data rather than relying on the operating system page cache. In bluestore you can adjust the amount of memory the OSD attempts to consume with the ``osd_memory_target`` configuration option. - Setting the osd_memory_target below 2GB is typically not recommended (it may fail to keep the memory that low and may also cause extremely slow performance. - Setting the memory target between 2GB and 4GB typically works but may result in degraded performance as metadata may be read from disk during IO unless the active data set is relatively small. - 4GB is the current default osd_memory_target size and was set that way to try and balance memory requirements and OSD performance for typical use cases. - Setting the osd_memory_target higher than 4GB may improve performance when there are many (small) objects or large (256GB/OSD or more) data sets being processed. .. important:: The OSD memory autotuning is "best effort". While the OSD may unmap memory to allow the kernel to reclaim it, there is no guarantee that the kernel will actually reclaim freed memory within any specific time frame. This is especially true in older versions of Ceph where transparent huge pages can prevent the kernel from reclaiming memory freed from fragmented huge pages. Modern versions of Ceph disable transparent huge pages at the application level to avoid this, though that still does not guarantee that the kernel will immediately reclaim unmapped memory. The OSD may still at times exceed it's memory target. We recommend budgeting around 20% extra memory on your system to prevent OSDs from going OOM during temporary spikes or due to any delay in reclaiming freed pages by the kernel. That value may be more or less than needed depending on the exact configuration of the system. When using the legacy FileStore backend, the page cache is used for caching data, so no tuning is normally needed, and the OSD memory consumption is generally related to the number of PGs per daemon in the system. Data Storage ============ Plan your data storage configuration carefully. There are significant cost and performance tradeoffs to consider when planning for data storage. Simultaneous OS operations, and simultaneous request for read and write operations from multiple daemons against a single drive can slow performance considerably. .. important:: Since Ceph has to write all data to the journal (or WAL+DB) before it can ACK writes, having this metadata and OSD performance in balance is really important! Hard Disk Drives ---------------- OSDs should have plenty of hard disk drive space for object data. We recommend a minimum hard disk drive size of 1 terabyte. Consider the cost-per-gigabyte advantage of larger disks. We recommend dividing the price of the hard disk drive by the number of gigabytes to arrive at a cost per gigabyte, because larger drives may have a significant impact on the cost-per-gigabyte. For example, a 1 terabyte hard disk priced at $75.00 has a cost of $0.07 per gigabyte (i.e., $75 / 1024 = 0.0732). By contrast, a 3 terabyte hard disk priced at $150.00 has a cost of $0.05 per gigabyte (i.e., $150 / 3072 = 0.0488). In the foregoing example, using the 1 terabyte disks would generally increase the cost per gigabyte by 40%--rendering your cluster substantially less cost efficient. .. tip:: Running multiple OSDs on a single SAS / SATA drive is **NOT** a good idea. NVMe drives, however, can achieve improved performance by being split into two more more OSDs. .. tip:: Running an OSD and a monitor or a metadata server on a single drive is also **NOT** a good idea. Storage drives are subject to limitations on seek time, access time, read and write times, as well as total throughput. These physical limitations affect overall system performance--especially during recovery. We recommend using a dedicated (ideally mirrored) drive for the operating system and software, and one drive for each Ceph OSD Daemon you run on the host (modulo NVMe above). Many "slow OSD" issues not attributable to hardware failure arise from running an operating system, multiple OSDs, and/or multiple journals on the same drive. Since the cost of troubleshooting performance issues on a small cluster likely exceeds the cost of the extra disk drives, you can optimize your cluster design planning by avoiding the temptation to overtax the OSD storage drives. You may run multiple Ceph OSD Daemons per SAS / SATA drive, but this will likely lead to resource contention and diminish the overall throughput. You may store a journal and object data on the same drive, but this may increase the time it takes to journal a write and ACK to the client. Ceph must write to the journal before it can ACK the write. Ceph best practices dictate that you should run operating systems, OSD data and OSD journals on separate drives. Solid State Drives ------------------ One opportunity for performance improvement is to use solid-state drives (SSDs) to reduce random access time and read latency while accelerating throughput. SSDs often cost more than 10x as much per gigabyte when compared to a hard disk drive, but SSDs often exhibit access times that are at least 100x faster than a hard disk drive. SSDs do not have moving mechanical parts so they are not necessarily subject to the same types of limitations as hard disk drives. SSDs do have significant limitations though. When evaluating SSDs, it is important to consider the performance of sequential reads and writes. An SSD that has 400MB/s sequential write throughput may have much better performance than an SSD with 120MB/s of sequential write throughput when storing multiple journals for multiple OSDs. .. important:: We recommend exploring the use of SSDs to improve performance. However, before making a significant investment in SSDs, we **strongly recommend** both reviewing the performance metrics of an SSD and testing the SSD in a test configuration to gauge performance. Since SSDs have no moving mechanical parts, it makes sense to use them in the areas of Ceph that do not use a lot of storage space (e.g., journals). Relatively inexpensive SSDs may appeal to your sense of economy. Use caution. Acceptable IOPS are not enough when selecting an SSD for use with Ceph. There are a few important performance considerations for journals and SSDs: - **Write-intensive semantics:** Journaling involves write-intensive semantics, so you should ensure that the SSD you choose to deploy will perform equal to or better than a hard disk drive when writing data. Inexpensive SSDs may introduce write latency even as they accelerate access time, because sometimes high performance hard drives can write as fast or faster than some of the more economical SSDs available on the market! - **Sequential Writes:** When you store multiple journals on an SSD you must consider the sequential write limitations of the SSD too, since they may be handling requests to write to multiple OSD journals simultaneously. - **Partition Alignment:** A common problem with SSD performance is that people like to partition drives as a best practice, but they often overlook proper partition alignment with SSDs, which can cause SSDs to transfer data much more slowly. Ensure that SSD partitions are properly aligned. SSDs have historically been cost prohibitive for object storage, though emerging QLC drives are closing the gap. HDD OSDs may see a significant performance improvement by offloading WAL+DB onto an SSD. One way Ceph accelerates CephFS file system performance is to segregate the storage of CephFS metadata from the storage of the CephFS file contents. Ceph provides a default ``metadata`` pool for CephFS metadata. You will never have to create a pool for CephFS metadata, but you can create a CRUSH map hierarchy for your CephFS metadata pool that points only to a host's SSD storage media. See :ref:`CRUSH Device Class` for details. Controllers ----------- Disk controllers (HBAs) can have a significant impact on write throughput. Carefully consider your selection to ensure that they do not create a performance bottleneck. Notably RAID-mode (IR) HBAs may exhibit higher latency than simpler "JBOD" (IT) mode HBAs, and the RAID SoC, write cache, and battery backup can substantially increase hardware and maintenance costs. Some RAID HBAs can be configured with an IT-mode "personality". .. tip:: The `Ceph blog`_ is often an excellent source of information on Ceph performance issues. See `Ceph Write Throughput 1`_ and `Ceph Write Throughput 2`_ for additional details. Additional Considerations ------------------------- You typically will run multiple OSDs per host, but you should ensure that the aggregate throughput of your OSD drives doesn't exceed the network bandwidth required to service a client's need to read or write data. You should also consider what percentage of the overall data the cluster stores on each host. If the percentage on a particular host is large and the host fails, it can lead to problems such as exceeding the ``full ratio``, which causes Ceph to halt operations as a safety precaution that prevents data loss. When you run multiple OSDs per host, you also need to ensure that the kernel is up to date. See `OS Recommendations`_ for notes on ``glibc`` and ``syncfs(2)`` to ensure that your hardware performs as expected when running multiple OSDs per host. Networks ======== Provision at least 10Gbps+ networking in your racks. Replicating 1TB of data across a 1Gbps network takes 3 hours, and 10TBs takes 30 hours! By contrast, with a 10Gbps network, the replication times would be 20 minutes and 1 hour respectively. In a petabyte-scale cluster, failure of an OSD drive is an expectation, not an exception. System administrators will appreciate PGs recovering from a ``degraded`` state to an ``active + clean`` state as rapidly as possible, with price / performance tradeoffs taken into consideration. Additionally, some deployment tools employ VLANs to make hardware and network cabling more manageable. VLANs using 802.1q protocol require VLAN-capable NICs and Switches. The added hardware expense may be offset by the operational cost savings for network setup and maintenance. When using VLANs to handle VM traffic between the cluster and compute stacks (e.g., OpenStack, CloudStack, etc.), there is additional value in using 10G Ethernet or better; 40Gb or 25/50/100 Gb networking as of 2020 is common for production clusters. Top-of-rack routers for each network also need to be able to communicate with spine routers that have even faster throughput, often 40Gbp/s or more. Your server hardware should have a Baseboard Management Controller (BMC). Administration and deployment tools may also use BMCs extensively, especially via IPMI or Redfish, so consider the cost/benefit tradeoff of an out-of-band network for administration. Hypervisor SSH access, VM image uploads, OS image installs, management sockets, etc. can impose significant loads on a network. Running three networks may seem like overkill, but each traffic path represents a potential capacity, throughput and/or performance bottleneck that you should carefully consider before deploying a large scale data cluster. Failure Domains =============== A failure domain is any failure that prevents access to one or more OSDs. That could be a stopped daemon on a host; a hard disk failure, an OS crash, a malfunctioning NIC, a failed power supply, a network outage, a power outage, and so forth. When planning out your hardware needs, you must balance the temptation to reduce costs by placing too many responsibilities into too few failure domains, and the added costs of isolating every potential failure domain. Minimum Hardware Recommendations ================================ Ceph can run on inexpensive commodity hardware. Small production clusters and development clusters can run successfully with modest hardware. +--------------+----------------+-----------------------------------------+ | Process | Criteria | Minimum Recommended | +==============+================+=========================================+ | ``ceph-osd`` | Processor | - 1 core minimum | | | | - 1 core per 200-500 MB/s | | | | - 1 core per 1000-3000 IOPS | | | | | | | | * Results are before replication. | | | | * Results may vary with different | | | | CPU models and Ceph features. | | | | (erasure coding, compression, etc) | | | | * ARM processors specifically may | | | | require additional cores. | | | | * Actual performance depends on many | | | | factors including drives, net, and | | | | client throughput and latency. | | | | Benchmarking is highly recommended. | | +----------------+-----------------------------------------+ | | RAM | - 4GB+ per daemon (more is better) | | | | - 2-4GB often functions (may be slow) | | | | - Less than 2GB not recommended | | +----------------+-----------------------------------------+ | | Volume Storage | 1x storage drive per daemon | | +----------------+-----------------------------------------+ | | DB/WAL | 1x SSD partition per daemon (optional) | | +----------------+-----------------------------------------+ | | Network | 1x 1GbE+ NICs (10GbE+ recommended) | +--------------+----------------+-----------------------------------------+ | ``ceph-mon`` | Processor | - 2 cores minimum | | +----------------+-----------------------------------------+ | | RAM | 24GB+ per daemon | | +----------------+-----------------------------------------+ | | Disk Space | 60 GB per daemon | | +----------------+-----------------------------------------+ | | Network | 1x 1GbE+ NICs | +--------------+----------------+-----------------------------------------+ | ``ceph-mds`` | Processor | - 2 cores minimum | | +----------------+-----------------------------------------+ | | RAM | 2GB+ per daemon | | +----------------+-----------------------------------------+ | | Disk Space | 1 MB per daemon | | +----------------+-----------------------------------------+ | | Network | 1x 1GbE+ NICs | +--------------+----------------+-----------------------------------------+ .. tip:: If you are running an OSD with a single disk, create a partition for your volume storage that is separate from the partition containing the OS. Generally, we recommend separate disks for the OS and the volume storage. .. _Ceph blog: https://ceph.com/community/blog/ .. _Ceph Write Throughput 1: http://ceph.com/community/ceph-performance-part-1-disk-controller-write-throughput/ .. _Ceph Write Throughput 2: http://ceph.com/community/ceph-performance-part-2-write-throughput-without-ssd-journals/ .. _Mapping Pools to Different Types of OSDs: ../../rados/operations/crush-map#placing-different-pools-on-different-osds .. _OS Recommendations: ../os-recommendations