I looked at the fast offline data to compare the HLT trigger efficiency and purity for vpdmb-5-p-sst events. Up to now, the hlt trigger is only used to label the events, so we have the full sample to compare the output from the HLT trigger vs. the regular trigger.

Data used are from fast offline production. Here shows a comparison of the refmult distributions between two triggers with offline |Vz|<6cm cut. (blue - vpdmb-5-p-sst, red - hlt)
The vertex was selected to be the vertex which satisfies |Vz-VzVpd|<3cm. The data shown here is from day 044. The right plot is a zoom-in version in the low refmult region.

One can see the HLT keeps all good events down to refmult ~40, and then starts to show inefficiency. Refmult~40 corresponds to roughly centrality ~ 60%. Down to refmult ~10 (which corresponds to ~80% centrality), the inefficiency is around 25%. Overall the average inefficiency at refmult>10 is about 97.5%. When we count all the events, the inefficiency is about 95.0%.

The following plot shows the Vz distributions between two triggers when refMult>10.
 
This shows also the online HLT vertex reconstruction purity is high. In this condition refmult>10, the purity to have real offline |Vz|<6cm is 99.6%.