The information in this thread is not correct.
In a scale out system you do get an XSEngine process on each node. The central web dispatcher uses a round robin load distribution approach to spread requests across each node.
Although I don't see where embedded mode is relevant here, that information stated previously also isn't correct. Embedded mode has been an option since SPS 05. However the default state remains non-embedded even on SPS 09 and beyond. The one exception is in MDC, each tenant DB instance has its own XSEngine process which does run embedded in an effort to keep the tenant footprint a little smaller.
As far as keeping XS from becoming a bottleneck: use it as intended. It's a lightweight pass thru layer. Keep your data intensive processing and flow within the DB by putting it in Views and Stored Procedures. Never return a result set to XS which is largwe than what you intend to send to the client side. Do all your filtering and paging within the data base. Remember that XS is always stateless.
