This article provides information on a known issue with SAP HANA and VMware VMXNET3 virtual NIC's.
Symptoms:
You may observe increased SAP HANA response time when a large number of concurrent users (~44,000) are running simultaneous transactions on the system.
See table below for latency and throughput numbers for vmxnet3 driver when compared against bare metal Linux installation:
Network Latency: VMXNET3 vs Bare Metal comparison (measured in microseconds with up too 65000 users)
| Baseline Latency with no load (us) | Latency at peak load with 44000-65000 concurrent users (us) | |
| Bare Metal | 26 | 95 |
| VMXNET3 | 84 | 318.82 |
Throughput: VMXNET3 vs Bare Metal comparison (measured in transactions per hour)
| 35% CPU 23000 Concurrent users | 65% CPU 44000 Concurrent users | Hi Load 64000 Concurrent users | |
| Bare Metal | 2914644 | 5550392 | 7105993 |
| VMXNET3 | 2911979 | 5502247 | 6472627 |
| Delta | -0.09% | -0.87% | -8.91% |
Notes:
Network latency of the VMXNET3 driver in µs and OLTP transaction response time in ms (SAP LUWs consist of many, for example hundreds of database LUW)
SAP and Database Logical Unit of Work (LUW):
A SAP LUW is a logical unit work that spans over several dialog steps. Let us consider a complex business transaction of n number of dialog steps; where each dialog step can have many DB LUW involving database operations that are logically related to each other. It is important that the changes to the database for this business transaction must be made all at once or it must be rolled back all for the data consistency; that is where the SAP LUW becomes critical in maintaining a logical relationship between many DB LUWs (dialog steps) & making all the changes to the database (COMMIT) in the final DB LUW.
The following SAP business transactions (interactive/online mode) were used to measure system performance :
In many cases these transactions will be run by users in online/interactive mode, however there are situations where the equivalent BAPI’s or IDOCS will be used via interfaces for batch or mass transactions. VMware recommends that users conduct performance tests on production like environments to determine if the increased response time will negatively impact their business processes.
VMware vSphere 7.0.x
VMware is actively investigating the root cause of this increased network latency. VMware is actively working with SAP to troubleshoot this issue and determine the appropriate corrective action.
This KB will be updated with the release version of the fix once the root cause is identified.