In this long overdue article, I would like to share an experience, where a customer was upgrading from Exchange 2010 to Exchange 2013. Note that this could also apply to customers migrating from Exchange 2007 or migrating to Exchange 2016 as well. The Exchange 2013 servers were hosted on VMWare vSphere 5.5U2; the Exchange 2010 servers on a previous product level.
The customer saw a negative impact on the end user experience of Outlook 2010 users, especially those working in Online Mode. Other web-based services like Exchange Web Services (EWS) were affected as well. The OWA experience was good.
After migrating end user mailboxes from Exchange 2010 to Exchange 2013 (but as indicated, this applies to Exchange 2016 as well), end users reported delays in their Outlook client responses, where sometimes Outlook seemed to ‘hang’ when performing certain actions like accessing a Shared Mailbox. Also, when opening up the meeting planner in order to schedule a room using Scheduling Assistant, it could take a significant amount of time, (i.e. minutes) before the schedule of all the rooms was being displayed.
The end users’ primary mailbox was configured to use Cached Mode, except for VDI users who used their primary mailbox in Online Mode. Shared Mailboxes were used in Online Mode due to the size (Outlook 2010, so no slider).
First, the overall health of the Exchange environment was checked to exclude it as a potential cause. Exchange performance metrics were monitored, as well as Managed Availability status and events, logs like the RCA logs, and VMWare CPU Ready % to check for potential vCPU allocation issues (read: oversubscription). None of these metrics caused any reason for concern.
After reconfiguring the HOSTS file, in order to bypass the load balancer and direct traffic to a single Exchange server to simplify troubleshooting, the symptoms remained. Then, we checked:
- TCP/IP optimization settings, e.g. RSS, Chimney, etc.
- VMWare VMXNet3 offloading, e.g. Large Send Offload, TCP Checksum Offloading
- VMWare VMXNet3 buffer settings
All those settings were also found to be on their recommended values.
We started digging in from the client’s perspective, and used WireShark to see what was going on on the wire. After filtering on the Exchange host, we saw the following pattern:
Note that this customer used SSL Offloading, so mailbox access took place on port 80 instead of 443 (RPC/http).
As you might notice, there is a consistent 200ms delay after the client receives its response (e.g. packets 106 and 110). When searching around for ‘200ms’ and ‘delay’, you may end up with articles describing the effect of the Nagle algorithm (Delayed ACK). Nagle is meant to reduce chatter on the wire, but can have a negative effect on near real-time communications, especially with small packets. Also, while 200ms might seem small, looking at the number of packets exchanged between Outlook and Exchange, this can add up quite quickly. Most of these articles will also describe a fix, recommending to configure a registry key TcpAckFrequency, and set it to 1 (default is 2). For testing purposes, we configured this key and after the mandatory reboot, the end user Outlook experience was snappy. However, setting this key would impact all client communications (real as well as VDI clients); not a recommended long-term solution due to side effects on the network.
After removing the registry key, investigating was continued. Since there was no issue with the Exchange 2010, we started to suspect there was perhaps an issue with VMWare, or there was some form of network optimization or packet inspection going on. This, due to the fact there was no problem with the old Exchange environment, and the elements that changed when migrating were VMWare vSphere version, physical vSphere hosts, and last but not least, the protocol switched. This client didn’t use Outlook Anywhere, so RPC/http was not enabled for Exchange 2010 prior to migration, and clients connected using MAPI. After some more investigating, some potentially related articles on the VMWare knowledgebase were found, talking about latency issues in certain VMWare Tools versions, the VMWare guest driver set, and downgrading these to 5.1 would have the same effect as configuring TcpAckFrequency. Unfortunately, this wasn’t an option as the hardware level of the VMWare guests already was on a certain level.
When installing VMWare Tools, the package comes with some system-level drivers which handle communications between the guest and the host or other guests. One of these drivers is the VMWare Guest Introspection driver (or VMCI Drivers, and formerly VShield Drivers). This component can be identified in the guest in the presence of the system drivers vnetflt and vsepflt, and accommodates agentless antivirus solutions like McAfee MOVE. However, it seems to also interfere with certain workloads in their driver ecosystem, thus negatively impacting real-time communications. I wasn’t able to test if the change from MAPI to RPC/http (or later MAPIhttp) also contributed to this effect, as the Introspection driver may not scan MAPI RPC packets at all, in which case there is no overhead introduced.
Needless to say disabling the Guest Introspection component might be less desirable for some organizations, and in those cases, when you experience this issue, I suggest contacting your VMWare representative, after verifying your VMWare Tools are part of the list of recommended versions.
In the end, in this situation Guest Introspection was disabled and a file-level scanner was introduced (with the required exclusions, of course). Performance for Online Mode was optimal when accessing Online Mode mailboxes, and using Exchange web services like Scheduling Assistant showed room planning in seconds rather than minutes.
Note that unfortunately, recent versions of VSphere running Exchange virtualized workloads also have this issue. On the plus side, they allow for separate (de)installation of the file system driver (NSX File Introspection Driver) and the network driver (NSX Network Introspection Driver). I am pretty sure removing the network driver would suffice, which might be a viable solution for some folks as well.
If you have any insights to share, please leave them in the comments.
Pingback: ohogy | IT?
Regarding changing the TcpAckFrequency registry key from 1 to 2, you state, “However, setting this key would impact all client communications (real as well as VDI clients); not a recommended long-term solution due to side effects on the network.” Can you be specific about the side effects?
Setting TcpAckFrequency to 1 disables Nagle, which results in acknowledging every packet, increasing the number of packets (SYN/ACKs without data), and thus generates overhead on the network. This may improve response times for applications requiring instant responses, but actual data throughput may be lower. And since it’s a system-wide setting, it may affect end user experience for all applications.
Do you know of any security implications of changing the TcpAckFrequency setting?
You actually make it seem really easy along with your presentation however I to find this matter to be actually something which I think I might never understand. It seems too complex and very vast for me. I am looking ahead in your next publish, I will attempt to get the grasp of it!
Facing the same issue plus outlook crashes in online mode in VDI, its not crashing in cached mode, tried removing guest introspection with no success same goes for tcpack.