 |
|
|
We used an IBM benchmark called "Advanced Workload Modeler
(AWM)". The benchmark simulates network traffic in a
given time frame. Adjustable parameters are packet size, runtime,
number of parallel connections and others.
An open source benchmark like iperf (http://dast.nlanr.net/Projects/Iperf/)
which provides less functionality can also be used to simulate
network traffic.
The following types of workload are simulated during our
tests:
Transactional workloads
With the request/response (rr) workloads connections to
the server are opened once. The workloads are then repeated
through the open connections for a fixed amount of measurement
time.
With the connect/request/response (crr) workload connections
to the server are opened and closed after completion for
each request to the server. The workload is repeated for
a fixed amount of measurement time.
Following is a list of transaction workloads:
- rr200x1000 (Send 200 bytes from client to server and
get 1000 bytes response) to simulate online transaction
requests
- rr200x32k (Send 200 bytes from client to server and
get 32k bytes response) to simulate database query requests
- crr64x8k (Open connection, send 64 bytes from client
to server and get 8k bytes response, close connection)
to simulate website requests
Streaming workloads
The streaming workload comprises two types. Strp stands
for "stream put" and strg stands for "stream
get".
Following is a list of streaming workloads:
- strp (Send 20Mbytes to the server and get 20 bytes response)
- strg (Send 20 bytes to the server and get 20m bytes response)
|
- IBM System z 2084_B16 (z990)
- OSA-Express (1GbE)
- OSA-Express2 (1Gigabit Ethernet, 1000Base-T, 10 Gigabit
Ethernet)
- 2 Linux LPARs for Linux to Linux measurements in LPAR
- 2 Linux guests under z/VM for Linux to Linux measurements
under z/VM
- Each Linux system with 4 CPUs, 2 GB memory
- z/VM 5.2
- Novell/SUSE SLES 10 and internal development kernel 2.6
- 5 minute workload with either 10 or 50 connections
|
This measurement was performed with an internal development
kernel 2.6 with two Linux LPARs.
The following chart shows the throughput of transactional
and file transfer workload for Gigabit Ethernet connections
via OSA-Express 1GbE and OSA-Express2 1GbE ports. The measurement
was performed via a crossed cable between two OSA CHPIDs with
10 simultaneous connections.
The x-axis names the transactional workloads. The file transfer
experiments strp and strg lead to identical results and are
shown only once.
OSA-Express2 offers up to 40% of throughput improvement over
OSA-Express.
|
|
This measurement was performed with an internal development
kernel 2.6 with two Linux LPARs.
The following chart shows the throughput of transactional
and file transfer workload for Gigabit Ethernet connections
via OSA-Express2 1 GbE and OSA-Express2 10 GbE ports. The
measurement was performed via a crossed cable between two
OSA CHPIDs with 50 simultaneous connections. The 50 simultaneous
connections have been selected here to generate the amount
of load that shows clearly the potential advantage of the
10 Gigabit Ethernet card over the 1 Gigabit Ethernet card.
The x-axis names the transactional workloads. The file transfer
experiments strp and strg lead to identical results and are
shown only once.
The larger the amount of transferred data in the requests,
the bigger is the throughput improvement with 10 Gigabit Ethernet.
With file transfer we reached a factor of 3.4x.
|
|
|
The measurements were performed with Novell SUSE SLES 10
with 2 Linux LPARs or with 2 Linux guests under z/VM 5.2.
As external connections we tested 3 OSA-Express2 cards,
- 1 Gigabit Ethernet
- 1000Base-T 1 Gigabit Ethernet
- 10 Gigabit Ethernet.
We measured the 2 MTU sizes 1492 and 8992.
For inter-LPAR connection we tested
- HiperSocket with MTU size 32k
and for inter-guest connections we tested
- HiperSocket
- GuestLAN type HiperSocket
- GuestLAN type QDIO
all with MTU size 32k.
The measurements were all done with 10 simultaneous connections.
The following charts show the throughput for all connection
types for the following workloads
|
The rr200x1000 shows with OSA-Express2 connections identical
throughput for the large and the default MTU size. The virtual
connections under z/VM provide almost double throughput, compared
to the OSA-Express2 numbers. Best results can be seen with
HiperSocket between 2 Linux LPARS.
|
|
The rr200x32k has larger amounts of data than the 200x1000
from the previous test. Here with OSA-Express2 connections
the larger MTU size shows a throughput advantage over the
default MTU size. The performance increase with the large
MTU size is more significant with the 10 GbE card (more than
2x). The virtual connections under z/VM show about triple
throughput, compared to the OSA-Express2 1 GbE numbers. Best
results are with HiperSocket between 2 Linux LPARS.
|
|
With the crr64x8k workload there is also an advantage for
the large MTU size with OSA-Express2 connections. The virtual
connections under z/VM have still double throughput of OSA-Express2
1 GbE. Best results are again with HiperSocket between 2 Linux
LPARS. The overall throughput advantage of the System z internal
connections over OSA-Express2 connections is here lowest among
the four workload types.
|
|
Strp and strg throughput is similar for all MTU sizes with
1 GbE OSA-Express2 cards. With the 10 GbE OSA-Express2 connection
with large MTU size has more than triple throughput of the
default MTU size. The virtual connections under z/VM have
more than triple throughput of the OSA-Express2 1 GbE cards.
Best results are again with HiperSocket between 2 Linux LPARS.
|
|
|
|
|
|
