Placement Group
- Strategy
- Cluster: low-latency in single AZ
- Spread: distinct hardware, minimize risk of failure
- Opposite of dedicated host
- You may get insufficient capacity when requesting
- Once provisioned you can stop and are guaranteed they will start
- Only some instance types are supported
- Instances may be moved in/out of Placement Group (since 2018)
- Works well with Enhanced Networking
- Use case
- Cassandra nodes
Enhanced Networking
- Traditionally EC2 uses split network driver model
- Frontend driver on the guest VM
- Backend driver on Xen Domain0
- Device driver - talks to Network device on physical host
- Data transferred through shared pages (ring buffer)
- Kernel 3.8.0+ has "grant pools" to speed up data transfer
- Alternatives
- Single Root I/O Virtualization
- Enhanced Network Adapter
Single Root I/O Virtualization (SR-IOV)
- PCIe device appear as many separate physical PCIe devices
- Physical Function (pf) - full featured PCIe functions (data,configuration)
- Virtual Functions (vf) - lightweight functions (data only)
- VM attaches to the vf
- Traffic goes directly to the device - does not traverse Domain0 at all
- i.e. NIC on instance driver talks to Device Driver
- OS must support it (
ixgbevfmodule) - Instance must support it (
sriovNetSupport)
Enhanced Network Adapter
- AWS version of SR-IOV
- Up to 25 Gbps
- Requires special driver on instance
EBS optimized instance
- Additional, dedicated capacity of EBS I/O
- Separated from normal network I/O traffic
- Depending on instance type:
- Throughput: 500 Mbps - 4000 Mbps, e.g:
- c4.xlarge = 500 Mbps
- c4.8xlarge = 4000 Mbps
- IOPS (16KiB): 450-3600
- Throughput: 500 Mbps - 4000 Mbps, e.g:
- Some instance types are EBS optimized by default (e.g. c4.large)
- Instance types with 10Gbps are different concept
- network channel is shared
- some do not offer EBS-optimized version at all (traffic shared)
- Example: i2.8xlarge
- Max
- Throughput: 800 MB/s (3200 Mbps)
- IOPS: 48,000
References
No comments:
Post a Comment