MayaScale Achieves 2.3M IOPS on Google Cloud N2 Instances

MayaScale's Ultra performance tier delivers a breakthrough on Google Cloud Platform: 2.3 million read IOPS with 173 microsecond latency, transforming N2 instances with local SSDs into ultra-low latency shared storage that rivals direct-attached NVMe performance.

Breaking the 2M IOPS Barrier on Google Cloud

Google Cloud's N2 instances with local SSDs have long been the go-to choice for performance-critical workloads. The n2-highcpu-64 instance provides 16 local NVMe drives with exceptional raw performance. However, until now, this performance came with a critical limitation: local storage isn't shared, and you lose all data if the instance terminates.

MayaScale solves this fundamental trade-off by pooling local NVMe storage across multiple instances using NVMe-over-Fabrics (NVMe-oF), delivering shared storage with latency that approaches—and in some cases matches—direct-attached NVMe performance. Our latest validation on GCP demonstrates performance that pushes the boundaries of what's possible with cloud storage.

Validated Performance Metrics

Latency Performance Curve

The latency performance curve demonstrates how MayaScale maintains sub-millisecond latency across different queue depths. Note that the best latency points (173μs read, 203μs write) both occur at QD8—showing optimal performance for moderately parallel workloads:

IOPS Scaling Performance

Queue depth 1 (QD1) performance is what matters for real applications—it's what your database, application server, or analytics engine actually experiences. But for workloads that can leverage parallelism, MayaScale's Ultra tier delivers extraordinary throughput. The performance curve below shows IOPS scaling with queue depth, with the sub-1ms latency zone highlighted:

Sequential Bandwidth Performance

In addition to exceptional random I/O performance, MayaScale delivers outstanding sequential throughput for large-block workloads such as data warehousing, backups, and streaming analytics:

Note: Write bandwidth is constrained by synchronous replication—every write must be committed to both storage nodes before acknowledgment. This ensures zero data loss during failover while still delivering 3.6 GB/s sustained throughput, which is more than sufficient for most sequential write workloads.

Why These Numbers Matter

The combination of ultra-low QD1 latency (192μs read) and massive parallel throughput (2.3M IOPS) makes MayaScale suitable for workloads previously confined to on-premises infrastructure or expensive specialized cloud instances:

• Single-threaded applications benefit from sub-200μs response times
• Highly parallel workloads can leverage millions of IOPS
• Database clusters get both low latency for individual queries and high throughput for concurrent operations
• AI/ML training experiences fast data loading at massive scale

Technical Architecture

Test Configuration

Our validation used Google Cloud's highest-performance instance configuration:

• Instance Type: n2-highcpu-64 (64 vCPU, 64 GB RAM)
• Local Storage: 16x local NVMe SSDs (375 GB each, 6.0 TB total)
• Network: 100 Gbps tier-1 networking
• Configuration: Active-Active with synchronous replication (dual-NIC architecture)
• Protocol: NVMe-over-TCP (NVMe-oF)
• Policy: zonal-ultra-performance
• Location: us-central1-f

Dual-NIC Architecture for Maximum Bandwidth

The Ultra tier uses a dedicated backend network (10.200.0.0/24) for server-side replication traffic, separating replication from client I/O. This dual-NIC design ensures that synchronous replication doesn't compete with client traffic, enabling both high availability and maximum performance.

This architecture is critical for write performance: with synchronous replication, every write operation must be committed to both storage nodes. By dedicating a separate 100 Gbps network path for replication, MayaScale achieves 866K write IOPS while maintaining sub-1ms latency.

How It Compares

vs. Google Cloud Persistent Disk

Metric	MayaScale Ultra	Persistent Disk Extreme	Difference
Read IOPS	2.3M	120K (max)	19x more IOPS
Read Latency	192μs	~1-2ms	5-10x lower latency
Usable Capacity	6 TB	User-defined	Fixed for Ultra tier
Configuration	2x n2-highcpu-64 32x local SSDs	Managed service	Self-managed instances
High Availability	Active-Active (synchronous replication)	Regional persistent disk available	Sub-second failover
Cost (monthly estimate)	~$5,500*	~$1,400**	Higher for extreme performance

*MayaScale Ultra: 2x n2-highcpu-64 with 16x 375GB local SSDs each, us-central1. Includes HA infrastructure but requires self-management.
**Persistent Disk Extreme: 6TB @ 120K IOPS (max tier), regional for HA. Managed service with lower operational overhead.
Note: MayaScale is designed for workloads where Persistent Disk's 120K IOPS limit is insufficient. If you need <120K IOPS, Persistent Disk is more cost-effective.

vs. Local SSD (Direct Attached)

MayaScale pools local SSDs to provide shared storage with near-local performance. While direct-attached local SSDs are slightly faster (no network overhead), they lack:

• Data persistence - Local SSDs are ephemeral; data lost on instance termination
• Sharing - Only one instance can access local SSDs
• High availability - No automatic failover

MayaScale delivers 80-90% of local SSD performance while adding enterprise-grade availability and shared access.

Perfect Use Cases for Ultra Tier

Getting Started

MayaScale Ultra tier can be deployed on Google Cloud Marketplace or via Terraform:

• Terraform deployment - Infrastructure-as-Code with full customization
• Policy-based selection - Specify "zonal-ultra-performance" policy for automatic instance selection
• Active-Active HA - Automatic failover with no downtime
• NVMe-oF client drivers - For Linux (RHEL, Ubuntu, Debian)
• CSI driver for GKE - Dynamic volume provisioning for Kubernetes
• Free trial available - Evaluate performance on your workload

Sample Terraform Configuration

module "mayascale_ultra" {
  source = "github.com/zettalane/terraform-gcp-mayascale"

  cluster_name        = "ultra-cluster"
  performance_policy  = "zonal-ultra-performance"
  zone                = "us-central1-f"
  project_id          = "your-project-id"

  # Automatically selects n2-highcpu-64 with 16 local SSDs
  # Deploys Active-Active with dual-NIC architecture
  # Target: 2.3M read IOPS, 866K write IOPS
}

Other Performance Tiers

Not every workload needs 2.3M IOPS. MayaScale offers five performance tiers on Google Cloud, from Basic (100K IOPS) to Ultra (2.3M IOPS). All tiers deliver sub-millisecond latency with Active-Active HA:

• Basic: 100K IOPS on n2-highcpu-4 - Development and testing
• Standard: 380K IOPS on n2-highcpu-8 - General purpose applications
• Medium: 700K IOPS on n2-highcpu-16 - Production databases
• High: 900K IOPS on n2-highcpu-32 - High-performance databases
• Ultra: 2.3M IOPS on n2-highcpu-64 - Maximum performance workloads

See our MayaScale on GCP page for detailed tier comparisons and performance graphs.

Why MayaScale on Google Cloud?

Conclusion

Google Cloud's N2 instances with local SSDs provide exceptional raw performance, but they've historically come with a trade-off: high performance or high availability, but not both. MayaScale eliminates this compromise.

With validated performance of 2.3 million read IOPS and 192 microsecond latency, MayaScale's Ultra tier on Google Cloud delivers shared storage that rivals direct-attached NVMe while providing enterprise-grade high availability. Applications that previously required on-premises infrastructure or expensive specialized hardware can now run efficiently on Google Cloud with performance that wasn't possible with traditional cloud storage.

The combination of Google Cloud's N2 hardware and MayaScale's software creates a compelling solution for databases, analytics, AI/ML training, and any workload that demands both extreme performance and high availability.