raid on Digi Hunch

High Performance Computing

Fri, 11 Dec 2020 23:42:00 -0400

Overview

High Performance Computing (HPC) has recently been commoditized with the advent of commodity server hardware (x86 server), virtualization technology and cloud delivery model. It is common in specialized industries where intensive computing tasks are required, for example:

HCL (healthcare and life science): drug discovery, computer aided diagnosis (CAD), genome engineering;
CAD, CAE, CAM (computer aided design, engineering, and manufacturing): 3D modeling, computational fluid dynamics (CFD), finite element analysis (FEA), structural mechanical design, etc
Finance: portfolio management, automated trading, risk analysis
Geoscience and geo-engineering: oil and gas exploration, geographic data, weather forecasting;
Scientific computation

Computing performance is measured in FLOPS (floating point operations per second) and is usually delivered in a cluster to aggregate the computing power from a number of networked nodes. This is referred to as an HPC cluster.

Hardware stack

An HPC cluster features the following components:

Head node (aka master node or login node): a gateway and coordinator; head node may be broken into several nodes
Compute node (worker node): the executor of jobs; the compute node can either be homogenous or heterogeneous, for different purposes. the number of compute nodes can be quite large

There are four common form factors for server: tower, rack-mount, blade, mainframe. Traditionally, the nodes are rack-mount 1U “pizza box” servers. Bladed systems started to replace due to the increased node density, thanks to the shared/redundant power and cooling management. In the past, the HPC cluster is operated in data centres, which is an expensive operation item. In the last decades, many organizations extends their compute workload to the cloud, forming a hybrid model.

HPC typically has specialized storage system because HPC applications notoriously create large amounts of data. NFS traditionally does not scale well as number of node increases. Some proprietary storage system such as Isilon provides good performance via NFS protocol. There are also open-source parallel file system such as Lustre and HDFS.

HPC networking handles three types of traffic:

computation traffic between compute nodes (if the compute nodes interact with each other)
file system traffic: for compute nodes to read and write on file system (e.g. NFS)
administrative traffic: fairly light compared to the two above

For that, many HPC runs two networks, a private (backend) network and a public (frontend) network. Backend network must be high speed and low latency, typically in the form of 10Gig Ethernet, or InfiniBand.

Software stack

On the software layer, the core functionality is Message Passing Interface (MPI), a specification for the developers and users of message passing libraries. MPI constitutes a standardized and portable message-passing system which consists of a library and a protocol to support parallel computing. MPI enables passing information between various nodes of a HPC cluster or between particular clusters, and has different implementations that provide the libraries to run HPC applications in a distributed manner across different physical nodes.

In the operation, user submits a job through head node in order to request the resource. User needs to specify the resources for the job (e.g. how many CPU cores, how much memory, etc). The head node runs a scheduler to allocate computing resource based on pre-defined policies, based on priority of jobs, availability of resources, distribution of load, etc. Depending on the nature of the computing jobs, the nodes participating in the task may or may not communicate with one another. If they do need to talk to each other, the program must support it. Such program can be called a cluster program, and the MPI (message passing interface) library greatly facilitates the development of such program. The sub-jobs communicating with each other also creates a considerable amount of network traffic within the cluster.

Cluster software ties all nodes in the cluster together. It turns raw hardware into a functioning cluster by provisioning (installing and configuring) the head nodes. Compute nodes can usually be added or removed dynamically therefore the head nodes should be able to provision compute nodes, and administer cluster, leaving the programming as the job for the user to complete. As mentioned, in parallel programming, the most important HPC tool is MPI (Message Passing Interface), which allows programs to talk to one another over cluster networks. There are both open (e.g. Open MPI) and commercial MPI (e.g. Microsoft MPI) versions. Cluster software should also provide compilers, debuggers, and profilers in addition to MPI.

There are cluster software in both Linux and Windows operating systems: Rocks Clusters, Oscar (Open Source Clusters Application Resources), Red Hat HPC solution, Microsoft HPC pack and AWS Parallel Cluster.

Implementation

Here is an example of setting up HPC cluster with CentOS. Despite of the well documented steps, note that the author of the document refers to HPC cluster simply as cluster, which is ambiguous. There are three basic motivators for creating a cluster: high performance computing (HPC), network traffic load balancing, and service resilience in the form of high availability (HA). The author should be specific in the document about the HPC cluster. If RDMA (Infiniband) network is involved, a configuration guide is provided in RedHat literature.

Here is an example of deploying HPC cluster in AWS. Here is the guide to deploy HPC pack in Microsoft technologies.

HPC and Big Data

HPC and Big Data are two distinctive computing paradigmes. Although there is some signs of convergence and blurred boundaries, it is still a long way before one can treat HPC and Big Data interchangeably. This paper does a phenomenal job in comparing the two paradigms.

The fundamental difference lies in the respective problems they intend to address. HPC focuses on the large computational loads, whereas Big Data targets applications that need to handle very large and complex data sets (usually in the order of multi-terabytes or exabytes). Many scientific data analytics applications are becoming I/O bound in modern systems, such as seismic algorithms, Big Data applications are thus very demanding in terms of storage, to accommodate such a masive amount of data, while HPC is usualy thought more in inters of sheer computational needs. The open-source projects in Big Data also aims to run on conventional hardware to make it easier and less expensive to scale. This is not the main focus of HPC.

So, you can run Big Data (e.g. Hadoop) analytics jobs on HPC gear. On the other hand, you can’t run HPC jobs on commodity hardware as commonly seen in the Big Data stack. Both HPC and Hadoop analytics use parallel processing of data. In a Hadoop/analytics environment, data is stored on commodity hardware and distributed across multiple nodes of hardware. In HPC, where the size of data file is much greater, data storage in centralized. Also, because of the sheer volume of its files, HPC also requires more expensive networking communications such as Infiniband, because the size of the file it processes require high throughput and low latency.

In BigData job, each query in Hadoop reads data from disk and runs as a separate MapReduce job. Spark enables in-memory iterative processing (through the RDD abstraction), allowing the user to query repeatedly on a dataset without having to perform intermediate disk operations. RDD are exposed in the Spark API where each dataset is represented as a read-only object, and transformations are invoked using methods on these objects. For an example project, check out this post.

The underlying software stacks for HPC and Big Data are fundamentally different, mainly due to the differences represent in their target class of applications, as outlined in the diagram below:

software stack difference between HPC and Big Data

As to which one is for me, the over-simplified advice is: if you can avoid HPC and just use Hadoop for your analytics, do it. It is cheaper, easier, and more cloud friendly. However, bear in mind that an all-Hadoop shop is not possible for many industries such as life sciences, weather, pharmaceutical, and academic applications.

Storage Nitty-Gritty 2 of 5 – SAN

Mon, 06 May 2019 22:06:02 -0400

In direct attached storage (DAS), storage is server centric and the host owns the storage. The storage is fully dedicated to the server that owns it.

With DAS, storage is server-centric

Storage has evolved into information centric model. In this model, when a new server is deployed in the environment, storage is assigned from the same shared storage pool to the new server.

Network based centralized storage solution

A network-based storage solution is the centralized storage pool. No single host owns the entire storage pool. The storage solution consists of two categories based on the interface with the host: SAN (storage area network) and NAS (network attached storage). To a client OS on the host, SAN typically appears as a local disk, allowing block-level access from the client OS, and therefore is more suited for structured workload such as database storage. It operates on its own storage network independent of the host network. NAS on the other hand, typically appears as a file share to the client OS, identified by an IP address and path. This is because NAS operates on the same TCP/IP network where the hosts are operated on. The client has file level access to storage, therefore NAS is better for unstructured data such as video and medical images. It is very important to understand the difference between SAN and NAS.

SAN deployment consists of two categories based on the connection technology. FC SAN is based on Fibre Channel network; and IP SAN is based on Internet protocol (iSCSI, FCIP, FCoE).

Fibre Channel SAN

Cable types: MMF (multimode fibre, usually for short distance within data centre because of signal attenuation due to modal dispersion) and SMF (single mode fibre, carries a single ray of light, used for long-distance cable runs;
Connector: standard connector (SC), lucent connector (LC) and straight tip connector (ST);
Interconnect device
- FC hub – for FC-AL implementation, but no longer in use
- FC switch – directly route data from one physical port to another (more intelligent than hub)
- Director – high end switches with redundant components to provide high availability
FC connectivity
- point-to-point: two devices connected directly to each other;
- arbitrated loop (FC-AL): devices are attached to a shared loop;

FC- AL (rarely used today)

switched fabric (FC-SW): uses switches that can switch data traffic between nodes directly through switch ports. Frames are routed between source and destination by the fabric

Fibre Channel Switched Fabric

Protocol: Fibre Channel Protocol (FCP): defines protocol stack (five layers, FC-0 through FC-4), addressing, identification (world wide name), frame, data structure, flow control, fabric services

Fibre Channel Protocol stack

FC frame

Topology: mesh topology and core-edge fabric topology
Block-level virtualization: aggregates block storage devices (LUNs) and enables provisioning of virtual storage volumes, independent of underlying physical storage. The virtualization layer maps the virtual volumes to the LUNs on the individual arrays. Block-level storage virtualization not only enables extending the storage volumes online; it consolidates heterogeneous storage arrays and enables transparent volume access. It also provides the advantage of non-disruptive data migration, where the virtualization layer handles the back-end migration of data, which enables the LUNs to remain online during migration.

Block-level virtualization (classic)

Federation of block storage across data centers (new generation)

Virtual SAN (VSAN, aka virtual fabric) – a logical fabric on an FC SAN, which enables communication among a group of nodes regardless of physical location in the fabric.

IP SAN

iSCI (one of the IP SAN protocols) – an IP based protocol that establishes and manages connections between host and storage over IP. iSCSI encapsulates SCSI commands and data into an IP packet and transport them using TCP/IP. It is relatively inexpensive and easy to implement so widespread in environments without FC SAN.

Topology
- Native connectivity (without FC components)
- Bridged connectivity (including FC components in the configuration)
- Combined connectivity (most common because a storage array usually comes with both FC and iSCSI ports)

iSCSI topologies

Protocol
- stack: SCSI is the command protocol that works at the application layer of OSI model; iSCSI is session-layer protocol that initiates a reliable session between devices that recognize SCSI commands and TCP/IP. The iSCSI session-layer interface is responsible for handling login, authentication, target discovery, and session management.

iSCSI protocol stack

iSCSI session and PDU encapsulation

PDU encapsulation

iSCSI discovery – an initiator must discover the location of its targets on the network and the names of the targets available to it before session establishment. Two types are SendTargets discovery and internet Storage Name Service
iSCSI names:
- IQN, iSCSI Qualified Name such as iqn.2008-02.com.example:optional_string;
- EUI, extended unique identifier such as eui.0300732A32598D26
iSCSI command sequencing – A command sequence may generate multiple PDUs. A command sequence number (CmdSN) within an iSCSI session is used for numbering all initiator-to-target command PDUs belonging to the session. This number ensures that every command is delivered in the same order in which it is transmitted, regardless of the TCP connection that carries the command in the session.

FCIP (one of the IP SAN protocols) – transporting FC block data over the IP infrastructure.

protocol stack and packet encapsulation

FCIP protocol stack

FCIP encapsulation

Topology (FCIP gateway involved):

FCIP topology

FCoE (one of the IP SAN protocols) – consolidation of LAN and SAN traffic over a single physical interface infrastructure. FCoE helps organizations address the challenges of having multiple discrete network infrastructures.

CNA (converged network adapters) replaces both HBAs and NICs in the server and consolidates both the IP and FC traffic
Special requirement on cables and switches
protocol stack and encapsulation

FCoE field mapping

CEE (converged enhanced Ethernet, or lossless Ethernet) provides new specification to existing Ethernet standard that eliminates the lossy nature of Ethernet. This makes 10Gb Ethernet a viable storage networking option, similar to FC. It features the following functionalities as part of IEEE 802.1:
- PFC (priority-based flow control)
- ETS (enhanced transmission selection)
- CN (congestion notification)

Storage Nitty-Gritty 1 of 5 – Disk and RAID

Sat, 09 Mar 2019 22:25:52 -0500

On my trip I have been through several pre-sales discussions on storage. Therefore I’m taking this opportunities to write up a series of postings deep diving into storage technologies. In this first section, we lay out the foundation of storage technology, from physical device to RAID, focusing on the concepts. Some contents are excerpts from Information Storage and Management.

Volume Manager – In early days, a file system occupies the entire disk drive, and presents continuous disk blocks directly to operating systems. Logical Volume Manager (LVM) was then introduced to bring a layer of abstraction (logical volume) on top of disks. The layers are shown as follows:

With all these layers, a byte in user file maps to disk sectors through several layers:

Mapping from user file to physical disk

File System – a hierarchical structure of files. It organizes data in a structural hierarchical manner. It includes files, directories as well as metadata. metadata must be consistent for the file system to be considered healthy. In Linux, metadata consists of:

Superblock: important information about file system, e.g. type, creation and modification dates, size, mount status flag
Inodes: a data structure that contains information associated with every file or directory
list of data blocks free and in use

Host connects to storage through various Interface Protocols. Common interface protocols include:

IDE/ATA and Serial ATA
SCSI (Small Computer System Interface)
FC (Fibre Channel)
IP (Internet Protocol per se is a network protocol traditionally used for host-to-host traffic in the early days. In the virtualization era, it has become a viable option for host-to-storage communication. Examples are iSCSI and FCIP)

The most prevalent disk drive types are SSD (solid state drive) and HDD (hard disk driveA). SSD (solid state drive) is newer, flash-based technology. Without seek and rotational latencies they deliver a high number of IOPS with low response times. They are especially suited for applications with small block size and random-read workloads requiring constant latency <1ms.

HDD is traditional and more cost effective. Its physical components are illustrated in the following two graphs:

Hard disk component

For HDD, disk service time (time taken by a disk to complete an I/O request) is determined by the following factors:

Seek time (aka access time) describes the time taken to position the R/W heads across the platter with a radial movement (moving along the radius of the platter). In other words, it is the time taken to position and settle the arm and the head over the correct track.
Rotation latency is the time taken by the platter to rotate and position the data under the R/W head. It depends on the rotation speed of the spindle and is measured in milliseconds.
(Data) transfer rate is the average amount of data per unit time that the drive can deliver from disk controller to the HBA (on the host).

Zone Bit Recording – a mechanism to use disk efficiently by grouping tracks into zones based on their distance from the disk.

Logical Block Addressing (LBA) – disk controller translates LBA to a physical address (CHS, cylinder, head and sector). The host only needs to know the size of disk drive in terms of number of blocks. The logical blocks are mapped to physical sectors on a 1:1 basis

IO request processing – I/O controller is introduce to improve response time for I/O request, in this model, The I/O requests arrive at the controller at the rate generated by the application. This rate is also called the arrival rate. These requests are held in the I/O queue, and the I/O controller processes them one by one, as shown here:

The relationship between controller utilization and average response time is:

Average response time = Service time / (1 – Utilization)

where service time is the time taken by disk controller to service the I/O request. This results in a classic relation between response time and utilization, as plotted below:

The graph indicates that the response time changes are nonlinear as the utilization increases. When the average queue sizes are low, the response time remains low. The response time increases slowly with added load on the queue and increases exponentially when the utilization exceeds 70 percent. Therefore, for performance-sensitive applications, it is common to utilize disks below their 70 percent of I/O serving capability.

Command queuing is a technique implemented on modern disk drives that determines the execution order of received I/Os and reduces unnecessary drive-head movements to improve disk performance. When an I/O is received for execution at the disk controller, the command queuing algorithms assign a tag that defines a sequence in which the commands should be executed. With command queuing, commands are executed based on the organization of data on the disk, regardless of the order in which the commands are received. Below is an example:

Disk Command Queuing

RAID is a technology that leverages multiple drives as part of a set that provides data protection against drive failures. It may also improve performance by serving I/Os from multiple disks simultaneously. It is primarily used in HDD but SSD may still benefit from it. RAID may be implemented by software but hardware RAID with a controller is widespread. RAID is built on three basic techniques:

Striping: spread data across multiple drives (more than one) to use the drives in parallel.
Mirroring: same data is stored on two different disk drives, yielding two copies of the data.
Parity: a method to protect striped data from disk drive failure without the cost of mirroring. An additional disk drive is added to hold parity, a mathematical construct that allows re-creation of the missing data.

Basic RAID levels are summarized here:

RAID summary

RAID 0, 1, 5 and 6 are pretty common in data centre operations. In addition to these levels above. If you hear RAID 1+0, 5+0 and RAID 6+0, they are called nested RAID. They are simply a RAID 0 on top of RAID1, RAID 5 and RAID 6, respectively.

RAID 0 – data striping technique utilizes full capacity of drives. Although it is a good option for applications that need high I/O throughput. It lacks data protection so it cannot drive application requiring high availability.

RAID 0

RAID 1 – mirroring technique ensures data duplication. In the event of disk failure, it introduces minimal impact to the disk array. It is suitable for applications that require high availability and cost is no constraint.

RAID 1

RAID 1+0 or RAID 10 – performs well for workloads with small, random, write-intensive I/Os. Some applications that benefit from RAID 1+0 include the following:

High transaction rate Online Transaction Processing (OLTP)
Large messaging installations
Database applications with write intensive random access workloads

RAID 1+0

RAID 3 – RAID 3 stripes data for performance and uses parity for fault tolerance. the total disk space required is 1.25 times the size of the data disks. RAID 3 always reads and writes complete stripes of data across all disks because the drives operate in parallel. RAID 3 provides good performance for applications that involve large sequential data access, such as data backup or video streaming.

RAID 5 – RAID5 is similar to RAID 4 because it uses striping. The drives (strips) are also independently accessible. The difference between RAID 4 and RAID 5 is the parity location. In RAID 4, parity is written to a dedicated drive, creating a write bottleneck for the parity disk. In RAID 5, parity is distributed across all disks to overcome the write bottleneck of a dedicated parity disk.

RAID 5

RAID 5 is good for random, read-intensive I/O applications and preferred for messaging, data mining, medium-performance media serving, and relational database management system (RDBMS) implementations, in which database administrators (DBAs) optimize data access.

RAID 6 – works the same way as RAID 5, except that RAID 6 includes a second parity element to enable survival if two disk failures occur in a RAID set. The write penalty in RAID 6 is more than that in RAID 5; therefore, RAID 5 writes perform better than RAID 6. The rebuild operation in RAID 6 may take longer than that in RAID 5 due to the presence of two parity sets.

RAID 6

Intelligent Storage System involves cache as the core component. An intelligent storage system involves front end, cache, back end and physical disks, as shown here:

Intelligent Storage System components

A variety of physical disk types and speed (e.g. mix of FC, SATA, SAS and flash) can be supported in a modern intelligent storage system. The front end provides the interface between the storage system and the host. It consists of ports and controllers, with redundancy. The back end provides an interface between cache and the physical disks. It consists of ports and controllers. For high data protection and high availability, storage systems are configured with dual controllers with multiple ports.

Cache improves storage system performance by isolating hosts from mechanical delays associated with hard disks. In intelligent storage system, read and write are first attempted on cache.

Page is the basic unit of cache, and the size of page is configured based on application I/O size. Cache consists of:

data store: holds the actual data temporarily
tag RAM: mainly serves three purposes:
- it tracks locations of data in cache and on disk;
- it maintains dirty bit flag to indicate whether data in cache has been committed to disk;
- it keeps time based information such as last access time, for cache management

Read Operation with Cache: When host issues a read request, the storage controller reads the tag RAM first to determine whether required data is available in cache:

Read cache hit: data is sent to host without any disk operation;
Read cache miss: back end access the disk to retrieve the requested data. Data is then placed in cache and sent to host through front end.

Cache miss increases I/O response time, to increase read hit ratio, read-ahead algorithm can be used when read requests are sequential. In a sequential read request, a contiguous set of associated blocks is retrieved. Several other blocks that have not yet been requested by the host can be read from the disk and placed into cache in advance. When the host subsequently requests these blocks, the read operations will be read hits.

Fixed prefetch – the intelligent storage system prefetches a fixed amount of data. It is most suitable when host I/O sizes are uniform.
Variable prefetch, the storage system prefetches an amount of data in multiples of the size of the host request.
Maximum prefetch limits the number of data blocks that can be prefetched to prevent the disks from being rendered busy with prefetch at the expense of other I/Os.

Write Operation with Cache: When an I/O is written to cache and acknowledged, it is completed in far less time (from the host’s perspective) than it would take to write directly to disk. Sequential writes also offer opportunities for optimization because many smaller writes can be coalesced for larger transfers to disk drives with the use of cache. Write operation with cache can be implemented in two ways:

Write-back cache: Data is placed in cache and an acknowledgment is sent to the host immediately. Later, data from several writes are committed (de-staged) to the disk. Write response times are much faster because the write operations are isolated from the mechanical delays of the disk. However, uncommitted data is at risk of loss if cache failures occur.
Write-through cache: Data is placed in the cache and immediately written to the disk, and an acknowledgment is sent to the host. Because data is committed to disk as it arrives, the risks of data loss are low, but the write-response time is longer because of the disk operations.

If the size of an I/O request exceeds the write aside size, writes are sent to the disk directly to reduce the impact of large writes consuming a large cache space. This is helpful where cache resources are constrained and cache is required for small random I/Os.

Cache space can be assigned in two ways:

dedicated cache: separate sets of locations are reserved for read and write;
global cache: user may specify percentage of cache for read and write based on application workload pattern; or the system set is dynamically.

Cache Management algorithm is used to determine when, and what pages of the cache need to be free up during maintenance. Most commonly used algorithms are:

LRU (least recently used): assuming data not accessed for a while will not be requested by host any more;
MRU (most recently used): assuming data recently accessed will not be requested by host again

As cache fills, the storage system must take action to flush dirty pages by committing data from cache to disk. There are several triggers for cache management action:

Idle flushing – occurs continuously at modest rate when cache utilization level is between high and low watermark;
High watermark flushing – activated when utilization hits high watermark; and stops at low watermark; this has impact to I/O processing;
Forced flushing – occurs in the event of large I/O burst when cache reaches 100% capacity; this significantly impacts I/O response time

Types of Flushing

Cache data protection is the mechanism to prevent losing uncommitted data held in cache. Common mechanisms are:

Cache mirroring – Each write to cache is held in two different memory locations on two independent memory cards. If a cache failure occurs, the write data will still be safe in the mirrored location and can be committed to the disk. The array operating environment needs to maintain cache coherency between the redundant memory locations. Read cache does not need mirroring.

Cache vaulting – In the event of server power failure, use battery power to write the cache content to the disk (vault drive). When power is restored, data from these disks is written back to write cache and then written to the intended disks.