mopdf.com

We designed Cisco HyperFlex systems as a next-generation platform that can adapt to meet newinformation technology challenges as they evolve. Cisco HyperFlex clusters are built with a minimum ofthree hyperconverged nodes and can support up to 64 nodes in a single cluster. (Limited-function edgeclusters can be built with as few as two hyperconverged nodes.) A significant differentiator for Cisco is ourCisco HyperFlex platform that support computing-only nodes, in a ratio of two computing-only nodes toone hyperconverged node. The computing-only nodes contribute CPU, memory, and graphics capabilitiesand share the storage presented by the Cisco HyperFlex hyperconverged nodes. The computing-onlynodes do not require the license costs for the Cisco HyperFlex HX Data Platform software. This factorpresents a significant economic advantage for VDI and SOLIDWORKS users, whose CPU and memorydemands outstrip storage requirements.The Cisco HyperFlex HX Data Platform is a purpose-built, high-performance, scale-out file system with awide array of enterprise-class data management services. The data platform’s innovations redefinedistributed storage technology, giving you complete hyperconvergence with enterprise storage features.Here is how the solution works.An HX Data Platform controller resides on each node and implements a distributed file system. Thecontroller runs in user space within a virtual machine and intercepts and handles all I/O from guest virtualmachines. Dedicated CPU cores and memory allow the controller to deliver consistent performance withoutaffecting the performance of the other virtual machines in the cluster. When nodes are configured withself-encrypting drives, the controller negotiates with Cisco UCS Manager to receive the encryption keysthat enable the drives to encrypt and decrypt data that flows to and from the various storage layers.The data platform has modules to support the specific hypervisor or container platform in use. Thecontroller accesses all of the node’s disk storage through hypervisor bypass mechanisms for excellentperformance. It uses the node’s memory and dedicated solid-state disk (SSD) drives or Non-VolatileMemory Express (NVMe) storage as part of a distributed caching layer, and it uses the node’s hard-diskdrives (HDDs), SSD drives, or NVMe storage for distributed storage. The data platform controller interfaceswith the hypervisor in two ways: IO Visor: The data platform controller intercepts all I/O requests and routes them to the nodesresponsible for storing or retrieving the blocks. IO Visor makes the existence of thehyperconvergence layer transparent to the hypervisor. Hypervisor agent: A module uses the hypervisor APIs to support advanced storage systemoperations such as snapshots and cloning. These are accessed through the hypervisor so that thehyperconvergence layer appears as if it were enterprise shared storage. The controller acceleratesoperations by manipulating metadata rather than performing actual data copying, providing rapidresponse and thus rapid deployment of new application environments.Data distributionThe HX Data Platform controller handles all read and write requests for volumes that the hypervisoraccesses and thus intermediates all I/O from the virtual machines and containers. Recognizing theimportance of data distribution, the HX Data Platform is designed to exploit low network latencies andparallelism, in contrast to other approaches that build on node-local affinity and can easily cause data hotspots.With data distribution, the data platform stripes data evenly across all nodes, with the number of datareplicas determined by the policies you set.This approach helps prevent both network and storage hot spots and makes I/O performance the sameregardless of virtual machine location. This feature gives you more flexibility in workload placement and 2021 Cisco and/or its affiliates. All rights reserved.Page 8 of 21

contrasts with other architectures in which a data locality approach does not fully utilize all availablenetworking and I/O resources. Data write operations: For write operations, data is written to the local SSD or NVMe cache, and thereplicas are written to remote caches in parallel before the write operation is acknowledged. Writeoperations are later synchronously flushed to the capacity layer HDDs (for hybrid nodes) or SSDdrives (for all-flash nodes) or NVMe storage (for NVMe nodes).Data read operations: For read operations in all-flash nodes, local and remote data is read directlyfrom storage in the distributed capacity layer. For read operations in hybrid configurations, data thatis local usually is read directly from the cache. This process allows the platform to use all solid-statestorage for read operations, reducing bottlenecks and delivering excellent performance.In addition, when a virtual machine migrates to a new location, the data platform does not require datamovement because any virtual machine can read its data from any location. Thus, moving virtual machineshas no performance impact or cost. The hyperconverged infrastructure is built using the building blocks listed in Table 3.Table 3.Hyperconverged infrastructure componentsProduct familyPlatformssupportedNVIDIA GPU supportCisco HyperFlex hyperconverged servers (hybrid, all flash,and all NVMe)1 rack unit (1RU)and 2RUNVIDIA T4, NVIDIA RTX 6000, andNVIDIA RTX 8000Cisco HyperFlex computing-only servers1RU, 2RU, andbladeNVIDIA P6, NVIDIA T4, NVIDIA RTX6000, and NVIDIA RTX 8000Cisco UCS 6154 and 64108 Fabric InterconnectsHCI andcomputing onlyCisco UCS Manager, Cisco UCS Central Software, andCisco Intersight platformHCI andcomputing onlyCisco Nexus 9300 platform Layer 2 switches (Cisco ACIsupported)HCI andcomputing onlyCisco security products (Cisco Duo, Cisco Umbrella, CiscoASA, Cisco Tetration Analytics, etc.)HCI andcomputing onlyCisco performance products (Cisco CWOM, CiscoAppDynamics, Cisco Stealthwatch Cloud, etc.)HCI andcomputing onlyAll NVIDIA Pascal, Turing, and RTXserver GPUsNVIDIA GPUs and software for virtualizationNVIDIA virtual GPU (vGPU) software enables delivery of graphics-rich virtual desktops and workstationsaccelerated by NVIDIA GPUs, the most powerful data center GPUs on the market today. With NVIDIA vGPUsoftware, GPU resources can be divided so that the GPUs are shared across multiple virtual machines, ormultiple GPUs can be allocated to a single virtual machine to power the most demanding workflo0ws.NVIDIA vGPU software runs on GPUs based on NVIDIA Turing, Volta, Pascal, Maxwell, and Ampere architectures.For SOLIDWORKS, four NVIDIA GPUs are best suited: 2021 Cisco and/or its affiliates. All rights reserved.Page 9 of 21

NVIDIA P6 GPU (Figure 3) Designed for blade servers and best suited for light users Delivers high graphics performance, improved energy efficiency, and increased user density with 16GB of frame buffer Mobile PCI Express (PCIe) module (MXM) form factor that runs on less than 90 watts (W) of power forhigh-density data centers Up to two NVIDIA P6 cards per Cisco UCS B200 M5 blade server.Figure 3.NVIDIA P6 GPU NVIDIA T4 GPU (Figure 4) With vWS software, it provides an excellent solution for light and medium 3D design and engineeringworkflows With a single slot, low-profile form factor, and just 70 W of power consumption, achieves maximumGPU density per server Up to six NVIDIA T4 cards per Cisco UCS C240 M5 or Cisco HyperFlex HX240c M5 rack server Up to two NVIDIA T4 cards per Cisco UCS C220 M5 or Cisco HyperFlex HX220c M5 rack server 2021 Cisco and/or its affiliates. All rights reserved.Page 10 of 21

Figure 4.NVIDIA T4 GPU NVIDIA RTX 6000 GPU (Figure 5) Combined with vWS software, enables design engineers to work from high-powered virtual designworkstations and render nodes to accelerate design workflows and arrive at their best creations faster With ray-tracing (RT) cores, a larger frame buffer, and multiple profile sizes, gives engineers anddesigners the flexibility to run demanding workloads from the data center Up to two NVIDIA RTX 6000 cards per Cisco UCS C240 M5 (July 2020) or Cisco HyperFlex HX240cM5 (Q4 2020) rack serverFigure 5.NVIDIA RTX 6000 GPU NVIDIA RTX 8000 GPU (Figure 6) Includes 48 GB of memory for the largest models—twice the frame buffer of the RTX 6000 Enables designers to work with the largest and most complex ray-tracing and visual-computingworkloads 2021 Cisco and/or its affiliates. All rights reserved.Page 11 of 21

Delivers the ultimate flexibility with vWS software, powering virtual design workstations and rendernodes to propel creative workflows Up to two NVIDIA RTX 8000 cards per Cisco UCS C240 M5 (July 2020) or Cisco HyperFlex HX240cM5 (Q4 2020) rack serverFigure 6.NVIDIA RTX 8000 GPUNVIDIA RTX Virtual Workstation (vWS) softwareFrom stunning industrial design to advanced special effects to complex scientific visualization, NVIDIA RTXis the world’s preeminent visual computing platform. And with NVIDIA RTX Virtual Workstation (vWS)software, you can now deliver the most powerful virtual workstation from the data center or cloud to anydevice, anywhere. Millions of creative and technical professionals can access the most demandingapplications from any device, work from anywhere, and tackle larger data sets, all while meeting the needfor greater security.Virtual workstations free users from the confines of physical location, delivering resources from the datacenter and providing secure access on any device, anywhere. NVIDIA vWS extends the trusted benefits ofNVIDIA RTX to deliver a true GPU-accelerated data center. This capability lets IT virtualize any applicationfrom the data center with a workstation-class user experience. Now your business can eliminateconstrained workflows that inhibit agility, and users can securely collaborate in real time without borders orlimits. You can efficiently centralize all your applications and data for dramatically lower IT operatingexpenses.Table 4 lists the main features of vWS software. These features allow excellent flexibility for supportingimportant capabilities and selecting the virtual workstation operating system. 2021 Cisco and/or its affiliates. All rights reserved.Page 12 of 21

Table 4.NVIDIA vWS featuresSizing considerationsYou need to consider four main factors when sizing a SOLIDWORKS deployment. Each is discussed here indetail. User types and requirements Display resolution and monitors CPU and memory selection GPU selection Scheduling engine Frame buffer and frame ratesUser types and requirementsGrouping users by their SOLIDWORKS use pattern is the first step in a successful virtual graphics workstationdeployment. Identify five main criteria for each group. Figure 7 shows the criteria. 2021 Cisco and/or its affiliates. All rights reserved.Page 13 of 21

Figure 7.User requirements User role: Identify the primary functions for SOLIDWORKS. Identify roles in terms of the intensity andpurpose of use. Complexity: Think in terms of read-only users, editors, designers, and the size of the system beingcreated or manipulated. Concurrency: How many applications does the user group member use at the same time? Are othergraphics-intensive applications besides SOLIDWORKS being used? Collaboration: Does the user group have a collaboration requirement that spans geographiclocations? Is the collaboration cross-functional? Across user roles? Working hours: Are your SOLIDWORKS users working across time zones? Does SOLIDWORKs useractivity follow the sun—that is, does user activity occur at all times?As a starting point, Cisco uses the classifications listed in Table 5.Table 5.Starting-point recommendations for user classificationsUser classificationCharacteristics Read-only, for review and documentationLight Project management Small parts or subassemblyMedium Read-only and full application Medium assembliesHeavy Full application Large assemblies or full productFor this document, virtual workstations will be constructed with the three user classifications listed in thetable. As stated previously, Cisco recommends running a proof-of-concept study with your users todevelop user classifications based on your specific use case.Display resolution and monitor selectionDifferent SOLIDWORKs user types generally have different requirements for display resolution and thenumber and types of monitors. This information is crucial in sizing the virtual graphic workstations that willsupport each user type. Resolution: HDMI, 4K, 5K, or higher Size: 24, 32, or 40 inches, or larger Number: 1, 2, or 4For this document, SPECviewperf 13 was used, with HDMI resolution, to measure performance anddetermine starting-point sizing recommendations for SOLIDWORKS.CPU and memory selectionNVIDIA GPUs work synergistically with the Intel Xeon Scalable family of processors and high-frequencymemory installed on the virtual graphics workstation host server. It is important to pair the proper CPU andmemory with the chosen NVIDIA GPU for the user type and virtual machine density expected for the virtualgraphic workstation hosts. 2021 Cisco and/or its affiliates. All rights reserved.Page 14 of 21

CPUs: High-frequency (3.0 GHz or faster) processors such as like the Intel Xeon Scalable 6248R,6246R, or 6242R are recommended. Memory: High-frequency (2933 MHz) memory in a balanced configuration (12 or 24 DIMMs) provideoptimal performance.GPU selectionEach of the user classifications in Table 5 may be served by a different NVIDIA graphics card. Carefulunderstanding of each user class’s use of the graphics card should be built by observation and bymeasurement using tools such as the GPUProfiler. In general, for SOLIDWORKs, Cisco recommends theNVIDIA T4, RTX 6000 and RTX 8000 (Table 6).Table 6.NVIDIA GPUs for SOLIDWORKS specificationsSpecificationsNVIDIA T4NVIDIA RTX 6000NVIDIA RTX 8000GPU and board(architecture)1 (Turing)1 (Turing)1 (Turing)Compute Unified DeviceArchitecture (CUDA) cores256046084608Tensor cores320576576RT cores407272Memory size16 GB GDDR624 GB GDDR648 GB GDDR6vGPU profiles1, 2, 4, 8, and 16 GB1, 2, 3, 4, 6, 8, 12, and 24 GB 1, 2, 3, 4, 6, 8, 12, 16, 24,and 48 GBForm factorPCIe 3.0 single slotPCIe 3.0 dual slotPCIe 3.0 dual slotPower70W250W250WThermalPassivePassivePassiveAs you can see in the table, all three cards are built on NVIDIA’s Touring architecture. Each card containsCUDA cores, Tensor cores, and ray-tracing cores. From a computing perspective, the T4 card is slightlymore than 50 percent of the RTX cards. The RTX cards have a larger processor and significantly morememory, making them excellent candidates for hosting very large models and rendering activities.Keep in mind that you can assign more than one NVIDIA GPU to a virtual workstation for demandingmultitaskers.The Cisco UCS C240 M5 server platform can accommodate up to six NVIDIA T4 cards or up to two NVIDIARTX cards. The Cisco UCS C220 M5 server can accommodate up to two NVIDIA T4 cards.Scheduling engine selectionNVIDIA offers a choice of three scheduling engines for use with SOLIDWORKS. Figure 8 provides details aboutthe choices available. 2021 Cisco and/or its affiliates. All rights reserved.Page 15 of 21

Figure 8.NVIDIA scheduling engine optionsAs shown in the figure, the best-effort scheduler is enabled by default. The idea behind this scheduler isthat all users assigned to a graphics card will seldom all use its capabilities simultaneously. This schedulerpotentially provides users with more GPU resources than their frame buffer allocation would suggest ifused strictly to allocate resources. This approach could result in inconsistent behavior if all users assignedto a card are actively using it. However, NVIDIA recommends using this method to schedule the GPU.Frame buffer selection and frame ratesFrame buffer selection is important for two reasons. First, it allocates the portion of the full GPU that will beallocated to the virtual machine that is assigned to it when all the virtual machines assigned to the card arepowered on and being actively used. The frame buffer cannot be oversubscribed. Second, the frame buffercan be used to allocate resources supported by GDDR6 memory. The larger the amount of frame bufferallocated, the more of a model that can be loaded into GPU memory for fast access and processing.The frame buffer is assigned through the NVIDIA profile selected when the virtual machine is created andthe NVIDIA PCI vGPU is assigned. There are several choices for the vWS profiles listed in Table 4.Frame rates are important for end users, particularly the heavy users identified in Figure 9. 2021 Cisco and/or its affiliates. All rights reserved.Page 16 of 21

Figure 9.Frame ratesTarget frame rates for smooth manipulation of SOLIDWORKS objects are 24 to 30 frames per second.Higher is better. If everything has been evaluated and provisioned correctly, all user types should achievethis performance level or better for their workloads.Keep in mind that for virtual graphic workstations, the maximum frame rate allowed by policy is 60 framesper second. The default policy for both Citrix Virtual Apps and Desktops and VMware Horizon is 30 framesper second.SOLIDWORKS sizing recommendationsCisco uses a tool from SPEC.org called SPECviewperf 13 to evaluate various GPUs, GPU profiles, and CPUcombinations to tune to starting-point recommendations for each individual. This is a publicly available toolthat you can use to conduct your own comparative testing. The tool can be found here.The SPECviewperf 13 benchmark is the worldwide standard for measuring graphics performance based onprofessional applications. The benchmark measures the 3D graphics performance of systems runningunder the OpenGL and Direct X APIs. The benchmark’s workloads, called viewsets, represent graphicscontent and behavior from actual applications.The latest version of the SPECviewperf benchmark is the SPECviewperf 13 benchmark, released on May23, 2018.The SPECviewperf 13 benchmark is a comprehensive upgrade of previous versions of the benchmark.Medical and energy viewsets incorporate new models and ray casting for volume visualization; the Mayaviewset features new models based on the SPECapc for Maya 2017 benchmark; and the Creo viewset hasbeen updated with fresh application traces. All other viewsets have been recompiled with minor changes.Results from the SPECviewperf 13 benchmark are not comparable to those from earlier versions.A SPECviewperf 13 benchmark license covers both Microsoft Windows and Linux versions; current paidlicense holders of the Windows version can receive the Linux edition free of charge.To evaluate SOLIDWORKS, SPECviewperf 13 was used in nonbenchmark mode, and that application onlywas evaluated with a series of graphic card, profile, and CPU and memory combinations. 2021 Cisco and/or its affiliates. All rights reserved.Page 17 of 21

NVIDIA RTX 6000 and RTX 8000 sample resultsFigures 10 through 13 show sample data collected from SPECviewperf 13 during the testing process forreference.Figure 10.8Q

Pure Storage FlashArray//m20, //m50, and //m70 arrays FlashStack Cisco security products (Cisco Duo, Cisco Umbrella , Cisco Adaptive Security Appliance [ASA], Cisco Tetration Analytics , etc.) FlashStack and FlexPod Cisco performance products (Cisco Workload Optimization Manager [CW