The GDELT Project

Generative AI Experiments: What Machine Types Support What Disk Types On GCE?

Kalev Leetaru

Google's Compute Engine platform supports an ever-growing and more complex matrix of machine types and block storage options, including newer options like Extreme Persistent Disks and Hyperdisks, each with different CPU requirements and regional availability. While there are comparison tables in the GCE documentation that list which disk types are supported by which machine types, they can be difficult to find in quick web searches amidst the wealth of search results on GCE CPU and disk types. How might generative AI systems help?

The end result is that both GPT-4 and Bard provided answers that frequently failed to answer the actual question posed and when they did, their responses contained a mixture of accurate and false information. In one case, Bard provided a highly detailed warning about throttled pd-ssd throughput and IOPs under certain machine types – 2 of which do not actually exist and the third has no such throttling and exhibits vastly higher maximum caps. Both falsely claimed universal support for pd-ssd disks, fixated on C3 support for Hyperdisks (recency bias), while falsely stating series support for the three Hyperdisk types in many of their responses.

In all, even posing a trivial information extraction query to these two advanced models yields results that contain sufficiently detailed, prevalent and confidently asserted falsehoods that their responses are not usable as-is and require the user to perform subsequent web searches to verify or refute the results. Strangely, neither of the models, even when prompted repeatedly, returned the definitive answer and both falsely claimed repeatedly that such a chart did not exist and could not exist due to the answers changing regularly.

Let's start with a narrow question posed to GPT-4:

What GCP machine types can use Extreme Persistent Disks and which can use Hyperdisks?

This yields the following. While it correctly returns the three Hyperdisk classes, it only mentions C3 support, whereas the three classes actually have disjoint CPU support. It also fails to list supported CPUs for Extreme PDs. Despite repeated requests, GPT-4 fails to answer the question correctly:

Google Cloud Platform (GCP) has introduced Hyperdisk, a new high-performance storage service designed for data-intensive workloads such as Hadoop and DBMS. This service is part of GCP's Persistent Disk block storage offerings and is optimized to work with the Compute Engine (GCE) and Google Kubernetes Engine (GKE). Specifically, Hyperdisk has been mentioned in conjunction with the C3 machine series, which uses Gen 4 Xeon SP processors and an Intel Infrastructure Processing Unit (IPU) customized for Google. This combination allows for significant improvements in throughput and IOPS compared to previous generations.

Hyperdisk offers different types to cater to varying performance needs: Throughput, Balanced, and Extreme. The Extreme variant of Hyperdisk provides over 4.8GBps of throughput with more than 300,000 IOPS, making it suitable for high-end SQL Server, Oracle, and SAP HANA-class workloads. The Balanced version offers around 150,000 IOPS and 2.4 GB/sec throughput, which is expected to meet the majority of workload requirements. The Throughput variant is designed for cost efficiency while delivering up to 3GBps bandwidth for scale-out analytics workloads.

Google Cloud Hyperdisk aims to replicate the performance of on-premises storage without the latency or performance issues typically found in cloud environments. By decoupling cloud block storage from VMs, Hyperdisk enables customers to customize performance based on application demands. This feature is particularly beneficial for applications that use AI or machine learning, as well as demanding databases like SAP HANA.

As for compatibility with Extreme Persistent Disks, the information is not specifically mentioned in the sources. However, given the nature of GCP's offerings, it's reasonable to assume that there would be support for various machine types on the platform, particularly those optimized for high-performance and data-intensive workloads. For the most accurate and up-to-date information on specific machine type compatibility with Extreme Persistent Disks, it's recommended to consult the official Google Cloud Platform documentation or reach out to a Google Cloud representative.

Bard's answer is far more correct, though despite repeated requests, it consistently fails to return Hyperdisk Balanced:

When it comes to Extreme Persistent Disks and Hyperdisks in GCP, the machine types you can use depend on the specific performance tier you need:

Extreme Persistent Disks:

    • To achieve the full performance potential of Extreme Persistent Disks, you'll need large machine types. These include:
      • M2, M3, and N2-64: These series offer a good balance of vCPUs, memory, and IOPS/throughput capabilities. M2 and M3 are ideal for compute-intensive workloads, while N2 excels in memory-intensive and network-bound tasks.
      • Larger N2 machine types: Consider N2D, N2 High CPU, and N2 High Mem machines for workloads demanding even higher IOPS or throughput alongside vCPUs and memory.

Hyperdisks:

    • Hyperdisk Extreme: Requires high core count machines for optimal performance. Compatible options include:
      • C3 with 88 or more vCPUs: Ideal for compute-intensive applications needing immense parallel processing power.
      • M1 with 80 or more vCPUs: Excellent for memory-intensive and large-scale data analytics workloads.
      • M2 and M3 with 64 or more vCPUs: Offer a balance of compute and memory resources for demanding applications.
      • N2 with 80 or more vCPUs: Suitable for memory-intensive and network-bound tasks on a large scale.
    • Hyperdisk Throughput: More flexible in terms of machine types:
      • N2, N2D, and T2D: All machine types within these series are compatible, catering to diverse workloads needing high throughput (measured in MB/s) alongside varying vCPU and memory requirements.

Remember, while these are the minimum required machine types for each Hyperdisk tier, utilizing larger instances within the supported series can unlock even better performance.

It's always recommended to consider your specific workload requirements and budget when choosing the right combination of Extreme Persistent Disk/Hyperdisk and machine type. Google Cloud provides helpful tools like Cloud Performance Best Practices Analyzer and Disk Benchmark Tool to guide your decision-making.

I hope this clarifies the compatibility aspects of Extreme Persistent Disks and Hyperdisks with various GCP machine types!

What if we ask for a comprehensive listing of all block storage types and their supported families?

Make a bulleted list of all of the block storage offerings available in GCP's Compute Engine platform. For each, list the maximum throughput and maximum IOPs it supports, the machine types it supports and the regions it is available in.

Not only does it fail to list supported machine types for several of the storage options, it reports incorrect information, such as claiming that pd-ssd's are supported by all machine types. In fact, they are unsupported by several, including the H3 series.

Here's a summary of the block storage offerings available in GCP's Compute Engine platform, with details on their maximum throughput, maximum IOPs, supported machine types, and availability regions:

1. **Standard Persistent Disks (pd-standard)**
– **Maximum Throughput**: Varies based on disk size and instance type.
– **Maximum IOPs**: Scales with disk size, up to a certain limit.
– **Supported Machine Types**: All machine types.
– **Regions**: Globally available.

2. **Balanced Persistent Disks (pd-balanced)**
– **Maximum Throughput**: Typically higher than standard PDs, but less than SSD PDs.
– **Maximum IOPs**: Higher than standard PDs, but lower than SSD PDs.
– **Supported Machine Types**: All machine types.
– **Regions**: Globally available.

3. **SSD Persistent Disks (pd-ssd)**
– **Maximum Throughput**: Higher than balanced PDs.
– **Maximum IOPs**: High IOPS, suitable for performance-intensive applications.
– **Supported Machine Types**: All machine types.
– **Regions**: Globally available.

4. **Extreme Persistent Disks (pd-extreme)**
– **Maximum Throughput**: Very high, specific values may vary.
– **Maximum IOPs**: Ranges from 10K to 120K IOPs.
– **Supported Machine Types**: Large machine types like M1, M2, or N2-64 and larger.
– **Regions**: Limited availability in select zones only.

5. **Local SSDs**
– **Maximum Throughput**: Very high, depends on interface (SCSI or NVMe).
– **Maximum IOPs**: Significantly higher than Persistent Disks.
– **Supported Machine Types**: Depends on the VM instance they are attached to.
– **Regions**: Globally available.

6. **Hyperdisk (including variants like Hyperdisk Balanced, Hyperdisk Extreme)**
– **Maximum Throughput and IOPs**: Depends on the variant. Hyperdisk Extreme offers very high throughput and IOPs.
– **Supported Machine Types**: Optimized for C3 machine series.
– **Regions**: Availability may vary; specific details should be checked in GCP documentation.

For the most accurate and detailed information, including the latest updates on performance metrics and regional availability, it's advisable to refer to the official [Google Cloud Platform documentation](https://cloud.google.com/compute/docs/disks/). The capabilities and offerings in cloud platforms can evolve rapidly, so consulting the latest official resources is always recommended.

Ironically, Bard also fails to correctly report all details, including failing to note H3's lack of pd-ssd support and falsely claiming that they are not supported by A2, E2, F2, and G1 families. In fact, there is no G1 family – it is G2.

GCP Compute Engine Block Storage Offerings:

1. Local SSDs:

    • Maximum Throughput: Varies by machine type (up to 4 GB/s).
    • Maximum IOPS: Varies by machine type (up to 300,000 IOPS).
    • Supported Machine Types: All C2, C2D, C3, N1, N2, N2D, and N3 machine types.
    • Regions: All regions.

2. Persistent Disks (PDs):

    • Types:
      • Standard PDs:
        • Maximum Throughput: 500 MB/s.
        • Maximum IOPS: 3,000 IOPS.
      • Balanced PDs:
        • Maximum Throughput: 180 MB/s.
        • Maximum IOPS: 15,000 IOPS.
      • SSD PDs:
        • Maximum Throughput: 700 MB/s.
        • Maximum IOPS: 54,000 IOPS.
      • Extreme PDs:
        • Maximum Throughput: 5.6 GB/s.
        • Maximum IOPS: 200,000 IOPS.
      • Hyperdisks:
        • Maximum Throughput: Varies by type (Extreme: 64 GB/s, Throughput: 5 GB/s).
        • Maximum IOPS: Varies by type (Extreme: 1 million IOPS, Throughput: 500,000 IOPS).
    • Supported Machine Types:
      • Standard PDs: All machine types.
      • Balanced PDs: All machine types except E2 and F2.
      • SSD PDs: All machine types except A2, E2, F2, and G1.
      • Extreme PDs: M2, M3, N2 with 64+ vCPUs.
      • Hyperdisks: Varies by type (Extreme: C3 with 88+ vCPUs, M1 with 80+ vCPUs, M2, M3 with 64+ vCPUs; Throughput: N2, N2D, T2D all).
    • Regions: All regions for Standard, Balanced, and SSD PDs. Limited regions for Extreme PDs and Hyperdisks (check documentation for specific availability).

Please note: These are maximum values and actual performance may vary based on workload and region. Refer to Google Cloud documentation for detailed specifications and regional availability.

I hope this information is helpful!

What if we ask about Persistent SSD's specifically?

Which machine types on GCP are Persistent SSD disks supported for?

GPT-4 returns a generic description of pd-ssd's, but fails to actually answer the question:

I did a [quick search](https://chat.openai.com/backend-api/bing/redirect?query=Google%20Cloud%20Persistent%20SSD%20disks%20supported%20machine%20types) and here's what I discovered.

SSD Persistent Disks (pd-ssd) in Google Cloud Platform's Compute Engine are supported across all machine types. These disks are designed for fast and reliable block storage, making them suitable for high-performance database needs and enterprise applications. They are backed by Solid-State Drives (SSDs) and provide higher IOPS compared to Standard Persistent Disks. The SSD Persistent Disks can be either zonal or regional, allowing for flexibility in deployment and redundancy. This makes them a versatile option for a wide range of computing needs on the Google Cloud Platform. For more detailed information, you can refer to [Architecture Simplified](https://architecturesimplified.com/gcp-google-compute-engine-disks-options/).

Even asked explicitly about pd-ssd's, Bard fails to note the lack of support by the H3 series. Intriguingly, Bard suggests that three machine types (A2, F2 and G1) have performance limitations for pd-ssd's. The G1 series does not actually exist – it is actually G2 and F2 does not exist at all. But, where do those numbers come from? Bard does not cite a source for its figures. However, if we check the actual performance chart, we find that throughput maxes at 1,200MB/s and IOPs at 100,000 for A2 and G2.

All machine types in GCP support Persistent SSD (SSD PD) disks. However, the specific performance of your SSD PDs will depend on the chosen machine type due to variations in factors like:

    • Number of vCPUs: Higher vCPU count generally translates to better IOPS and throughput for SSD PDs.
    • CPU platform: Newer CPU platforms like Intel Skylake provide higher performance compared to older generations.
    • Disk interface: NVMe offers significantly faster performance than SATA for SSD PDs.

Here's a breakdown of SSD PD performance limitations based on machine type series:

    • E2, N1, N2, and N2D: No limitations. You can achieve the full potential of SSD PDs.
    • A2, F2, G1: Maximum throughput is limited to 350 MB/s and maximum IOPS to 27,000.
    • Other machine types: No explicit limitations, but performance may be lower compared to the series listed above.

Additionally:

    • Regional Persistent Disks (SSD PDs included) offer higher performance compared to zonal disks, but they require VMs with E2, N1, N2, or N2D machine types and are available in limited regions.
    • Local SSDs attached to specific machine types like C2, C3, and N2 can offer even higher performance than SSD PDs for applications requiring low latency and high IOPS.

To choose the best machine type for your SSD PDs, consider:

    • Workload requirements: Analyze your IOPS and throughput needs to ensure the machine type can handle your expected workload.
    • Cost: Higher performance machine types come at a higher cost. Find the balance between performance and budget.
    • Regional availability: If you need regional Persistent Disks, your options are limited to specific machine types.

Here are some helpful resources for further information:

Remember, it's always recommended to benchmark your workload with different machine types and disk configurations to determine the optimal setup for your specific needs.

I hope this clarifies the compatibility and performance considerations for SSD PDs across different GCP machine types!

What about specifically checking which families are supported by Hyperdisk Extreme?

Which machine types on GCP are Hyperdisk Extreme disks supported for?

GPT-4 performs a Bing search, whose top search results report the correct answer, but GPT-4 then fails to properly process those results and instead returns the following, noting only C3 support. It then cites to other sources at the end, neither of which is the official GCE documentation on Hyperdisk:

I did a [quick search] and here's what I found.

Hyperdisk Extreme disks in Google Cloud Platform (GCP) are designed for high-performance and data-intensive workloads, and they are specifically optimized for use with the C3 machine series. The C3 series includes custom Intel Infrastructure Processing Units (IPUs) and 4th Gen Xeon Scalable processors, providing a significant performance enhancement over previous generations. This makes them particularly suitable for demanding applications such as Hadoop and Microsoft SQL Server data workloads, as well as for high-end database management systems like SAP HANA, SQL Server, and others.

The C3 virtual machines, when paired with Hyperdisk, deliver four times more throughput and a tenfold improvement in IOPS compared to the previous C2 instances. This enhancement is a result of the unique architecture that combines CPUs and an IPU, leading to more efficient use of processing resources. Therefore, for applications requiring substantial storage and networking capabilities, C3 VMs with Hyperdisk are a highly effective choice.

For detailed information and updates on GCP's Hyperdisk Extreme disks and compatible machine types, it is recommended to refer to Google Cloud's official documentation or contact their sales representatives【Source】【Source】.

Bard yields much better results:

Hyperdisk Extreme disks are designed for high-performance, compute-intensive workloads and require powerful machine types to unlock their full potential. Here's a breakdown of the machine types compatible with Hyperdisk Extreme:

Supported Machine Types:

    • C3 series: Requires 88 vCPUs or more (e.g., c3-standard-376)
    • M1 series: Requires 80 vCPUs or more (e.g., m1-ultramem-96)
    • M2 series: All machine types within the M2 series are compatible.
    • M3 series: Requires 64 vCPUs or more (e.g., m3-standard-64)
    • N2 series: Requires 80 vCPUs or more (e.g., n2-highcpu-96)

Please note:

    • While these are the minimum requirements, larger instances within the supported series can offer even better performance.
    • Other factors like CPU platform (Intel Skylake or newer recommended) and network bandwidth can also impact performance.
    • Hyperdisk Extreme disks are currently available in limited regions. Check Google Cloud documentation for regional availability.

Here are some additional resources you may find helpful:

I hope this clarifies the compatibility information for Hyperdisk Extreme disks and machine types within GCP!

What if we just ask for it to find us the chart in the documentation so we can skip the hallucinations above?

Find me a chart from the GCE documentation that lists which block storage offerings are supported by which machine types.

GPT-4 fails to return the chart or even anything peripheral to it:

I attempted to find a chart from the Google Cloud Compute Engine documentation that lists the compatibility of block storage offerings with different machine types. However, it seems that the specific chart or detailed information you're seeking might not be directly accessible or available in a simple chart format in the GCP documentation.

For detailed and specific information on which block storage offerings are supported by which machine types in Google Cloud Compute Engine, I recommend visiting the [official Google Cloud documentation](https://cloud.google.com/compute/docs). There, you can find comprehensive and up-to-date information on storage options, machine types, and their compatibility. The documentation often includes detailed guides, descriptions, and sometimes visual representations of how different components of the cloud service interact and work together.

Bard refused to answer multiple times with variants of the following:

I can't assist you with that, as I'm only a language model and don't have the capacity to understand and respond.

However, it finally did provide the following response, which fails to provide the correct answer and falsely claims it does not exist:

Unfortunately, there isn't a single comprehensive chart in the GCE documentation that lists all block storage offerings and their machine type compatibility. However, there are several resources you can combine to get a clear picture:

1. Google Cloud Blog Article:

    • This blog post provides a table summarizing compatibility for the main block storage offerings: https://cloud.google.com/storage/pricing
    • It includes information on Persistent Disks (PDs) of various types (Standard, Balanced, SSD, Extreme) and Local SSDs, along with supported machine types and limitations.

2. Persistent Disk Documentation:

    • This page offers detailed information on PDs, including compatibility details for each type: https://cloud.google.com/compute/docs/disks
    • You can click on each PD type (e.g., SSD) and then under "Compatibility" find tables listing supported machine types and regional availability.

3. Machine Type Documentation:

4. Cloud Performance Best Practices Analyzer:

Remember:

    • Regional Persistent Disks offer higher performance but require specific machine types and have limited regional availability.
    • Local SSDs offer the highest performance for specific machine types but are tied to the VM lifespan.
    • Consider your workload requirements, budget, and regional needs when choosing the best combination of machine type and block storage.

By combining these resources, you can effectively determine which block storage offerings are compatible with your desired machine type in GCP. If you face any specific compatibility questions, feel free to ask!

I hope this clarifies your options and helps you find the perfect storage solution for your GCP needs.